# Building 0 > Student-run think tank and prototyping lab at SUTD (DAI Fab Lab). We investigate how emerging technology reshapes systems, institutions, and the built environment — through building, testing, and iteration. The site is a JavaScript application. Every page below also has a plain-text Markdown mirror that requires no JavaScript. Prefer the `.md` links when reading programmatically, or fetch `/llms-full.txt` for everything at once. - Full text of the whole site: https://bldg0.com/llms-full.txt - Sitemap: https://bldg0.com/sitemap.xml ## Pages - [Building 0 — Student-run think tank & prototyping lab at SUTD](https://bldg0.com/): Lab identity, position, and operating philosophy of Building 0, a student-run think tank and prototyping lab at SUTD (DAI Fab Lab). Plain text: https://bldg0.com/index.md - [About Building 0](https://bldg0.com/about): What Building 0 is, why it exists, and how it operates. Plain text: https://bldg0.com/about.md - [Projects — Building 0 Cohort Snapshot](https://bldg0.com/work): Cohort project index, impact matrix, recognition, media coverage, and stakeholder engagement. Plain text: https://bldg0.com/work.md - [Blog — Building 0](https://bldg0.com/writing): Index of essays, papers, and field notes from active explorations. Plain text: https://bldg0.com/writing.md - [Join Building 0](https://bldg0.com/join): How to join the lab, who thrives here, and what the first term looks like. Applications go to anieyrudh_r@mymail.sutd.edu.sg. Plain text: https://bldg0.com/join.md ## Blog — essays, papers and field notes - [The Dream Engine](https://bldg0.com/writing/the-dream-engine): On AI, taste, and why creation requires risk. — Essay, Chen Liang Jung, 14 Apr 2026. Plain text: https://bldg0.com/writing/the-dream-engine.md - Sections: I · The Entropy Artist · II · Taste as Taming · III · The Tracing Paper · IV · The Unfinished Pedagogy - [Intelligence in Deployment](https://bldg0.com/writing/intelligence-in-deployment): Most AI systems that appear intelligent are not. They perform without closing the loop — and that distinction is the one that matters in the real world. — Paper, Anieyrudh R, 8 Apr 2026. Plain text: https://bldg0.com/writing/intelligence-in-deployment.md - Sections: Core Thesis · The Loop · The Threshold Claim · The Diagnostic Structure · The Evaluation Principle · Domain Application · Deep Walkthrough: Robotics · Positioning Against Literature · References - [The New Nature of Work](https://bldg0.com/writing/the-new-nature-of-work): Why AI Dissolves the Contract and What Takes Its Place — Essay, Anieyrudh R, 29 Mar 2026. Plain text: https://bldg0.com/writing/the-new-nature-of-work.md - Sections: The Question Nobody Is Asking · Two Definitions of Work · What Cannot Be Automated · Why This Requires Presence · The Redesign Required - [Agentic Presence](https://bldg0.com/writing/agentic-presence): Maintaining Intellectual Agency in AI‑Mediated Learning — Paper, Anieyrudh R, 22 Nov 2025. Plain text: https://bldg0.com/writing/agentic-presence.md PDF: https://bldg0.com/papers/agentic-presence.pdf - Sections: Abstract · 1. Introduction · 2. Why This Matters · 3. Construct Definition · 4. Intentionality · 5. Discernment · 6. Ownership · 7. Integration · 8. Capability Types · 9. Design Environments · 10. Open Questions · 11. Conclusion · References · Appendix A ## Project write-ups - [Don't Think, Just Ask](https://bldg0.com/work/dont-think-just-ask): SUTD Wiki: building an AI chatbot that became a diagnostic for institutional knowledge failure. Plain text: https://bldg0.com/work/dont-think-just-ask.md ## Projects (33 total) - TextForm — Wang Kangjie Bryan. Architecture & design. Deployed, Validated. https://github.com/WKJBryan/Text-to-3D - SUTD Wiki — Anieyrudh R. Learning & education. Scaled, Institutionalized. https://bldg0.com/work/dont-think-just-ask - Media Release GPTs — Anieyrudh R, with Yvonne (Office of Marketing). Professional workflows. Deployed, Measurable. - EJC AI Fluency Programme — Anieyrudh R (lead); Bryan, Liang Jung, Joshua, wider B0 team. Learning & education. Scaled, Institutionalized. https://www.sutd.edu.sg/wp-content/uploads/2026/04/Designing-A-Human-AI-Innovation-Edge-For-Enterprises-April-2026.pdf - EAB Copilot Agent — Anieyrudh R, for Mr Toe Oo Zaw (Senior Manager, Accreditation). Learning & education. Deployed, Measurable. - Wallee — Anieyrudh R, Puah Joon Leng. Architecture & design. Deployed, Validated. https://github.com/anieyrudh/Wallee - JTC × MUSPP Knowledge Graph — Chen Liang Jung (lead), Wang Kangjie Bryan. Government / public sector. Scaled, Validated. - Grafyn — Wang Kangjie Bryan. Professional workflows. Deployed, Validated. https://github.com/WKJBryan/Grafyn - Social Matrix — Wang Kangjie Bryan. Professional workflows. Deployed, Validated. - AlmondMCP — Chen Liang Jung. Architecture & design. Deployed, Validated. - Betelnut — Chen Liang Jung. Government / public sector. Deployed, Validated. - Chestnut — Chen Liang Jung. Architecture & design. Deployed, Validated. - cadquery-library — Wang Kangjie Bryan. Architecture & design. Deployed, Measurable. https://github.com/WKJBryan/cadquery-library - SO101 Physical Set-Up Planner — Puah Joon Leng (lead), Anieyrudh R. Architecture & design. Deployed, Measurable. https://so101-camera-planner.netlify.app - SUTD Deadline Centre — Abhishek Vulla. Learning & education. Deployed, Measurable. https://github.com/AbhishekVulla/sutd-deadline-center - LearnLoop — Abhishek Vulla. Learning & education. Pilot, Measurable. - Planspear — Bervin Bek (external partner). Government / public sector. Pilot, Emerging. - PA Volunteer Mgmt (Chingay) — Anieyrudh R, for People's Association. Government / public sector. Prototype, Emerging. - SUTD Atlas — Puah Joon Leng (lead), Anieyrudh R. Learning & education. Prototype, Emerging. - Paideia — Anieyrudh R (directing). Learning & education. Prototype, Exploratory. https://github.com/anieyrudh/Paideia - clicky-windows — Abhishek Vulla. Learning & education. Prototype, Exploratory. https://github.com/AbhishekVulla/clicky-windows - AutoFlowMatic AI Sales Agent — Stephen Huang (external partner). Professional workflows. Prototype, Emerging. - DeepLearning — Abhishek Vulla. Professional workflows. Prototype, Exploratory. - Truck Load Planner — Abhishek Vulla. Professional workflows. Prototype, Emerging. - Risk Assessor AI — Poh Qiao Hui Stefanie (ex-member). Professional workflows. Prototype, Emerging. - Gen AI OCR (SUTD Finance) — Anieyrudh R, for SUTD Finance. Professional workflows. Prototype, Emerging. - Loom — Anieyrudh R. Professional workflows. Prototype, Exploratory. - Brinepool — Cheng Hui Joshua. Professional workflows. Pilot, Emerging. https://brinepool.vercel.app - Manpower Mgmt (Shangri-La) — Anieyrudh R, for Shangri-La Singapore. Professional workflows. Pilot, Emerging. - Whisper Library — Poh Qiao Hui Stefanie (ex-member). Professional workflows. Pilot, Emerging. https://whisperlibrary.up.railway.app - ITE × SUTD D.A.R.E. — Anieyrudh R (co-lead organiser & lead technical instructor). Learning & education. Scaled, Institutionalized. - Corpus — Anieyrudh R. Professional workflows. Deployed, Validated. https://corpus-memory-sin-425b92.fly.dev/guide - B0 Update Tracker — Abhishek Vulla. Professional workflows. Deployed, Measurable. https://b0-update-tracker.up.railway.app ## Contact - Applications and enquiries: anieyrudh_r@mymail.sutd.edu.sg --- [Image: Building 0 — fabrication lab with robotics, prototypes, and active workstations] Student-run think tank & prototyping lab # Build the Future of Building 0 is a student-run think tank and prototyping lab. We investigate how emerging technology reshapes systems, institutions, and the built environment. We build to find out what holds. P.01PRINCIPLE In an age of abundance, constraint is a design choice. When tools become abundant, discernment matters more than access. [Image: Prototyping workspace — material studies] Fig. 01 Active workspace — material prototyping + structural testing 02 — Position ## We don't optimize. We redesign the question. We work across design, engineering, and institutional thinking — prototyping futures, not products, in domains where technology creates structural possibility. [About the Lab →](/about) P.04PRINCIPLE We build to find out what holds. Prototypes are one way to think through uncertainty. 04 — How We Work [Read More →](/about) 01 ### High Autonomy Define your own problems. Direction follows evidence, not hierarchy. 02 ### Rapid Iteration Build-test-learn in days, not quarters. Prototypes are how we think. 03 ### First-Principles Framing Every project starts by questioning the assumptions behind the question. 04 ### Cross-Functional by Default Research, design, engineering, and systems thinking converge in every project. 05 ### Show, Don't Tell Working artifacts over pitch decks. Tested prototypes over strategy memos. 06 ### Technology as Means AI and computation are levers — not identities. The question comes first. 05 — Open ## Work on problems that matter. Build things that get tested. Building 0 is looking for people who want real ownership of open-ended, consequential work — and collaborators with live problems worth exploring. [Join Building 0 →](/join)[Start a Conversation →](/join) --- 00 — About # A specific way of thinking, working, and building. 01 — What Building 0 Is Building 0 is a student-run think tank and prototyping lab at SUTD, also known as the DAI Fab Lab. We explore the Future of X — how emerging technology, institutional shifts, and new systems reshape the world. We work across design, AI, engineering, and systems thinking. As a high-autonomy lab, we engage industry partners, public agencies, and the wider SUTD community to investigate problems without settled answers — building prototypes, pressure-testing ideas, and iterating at the pace of real change. Think TankPrototyping LabCross-DomainHigh Autonomy [Image: Building 0 lab space — active prototyping environment] Fig. 01 — Lab interior P.04PRINCIPLE We build to find out what holds. Prototypes are one way to think through uncertainty. 02 — Why It Exists The most consequential questions about technology and institutions do not come with ready-made answers. Academic timelines are too slow. Industry incentives are too narrow. Policy lags behind technical reality. Building 0 exists to investigate those questions through building, testing, and iteration — not commentary alone. Working prototypes, tested systems, and documented evidence of what works and what does not. 03 — How It Operates 01 ### High Autonomy Define your own problems. Direction follows evidence, not hierarchy. 02 ### Rapid Iteration Build-test-learn in days, not quarters. Prototypes are how we think. 03 ### First-Principles Framing Every project starts by questioning the assumptions behind the question. 04 ### Cross-Functional by Default Research, design, engineering, and systems thinking converge in every serious project. 05 ### Show, Don't Tell Working artifacts over pitch decks. Tested prototypes over strategy memos. 06 ### Technology as Means AI and computation are levers — not identities. The question comes first. 04 — The Environment The lab is intense but generative. Problems shift as evidence comes in. People take ownership of their domains and share findings openly — including failures. You may be researching in the morning, prototyping in the afternoon, and rewriting the framing after new evidence arrives. Collaboration is Bayesian, low-ego, and oriented toward learning speed. PaceDays, not quarters OwnershipDefine, build, test, document FeedbackDirect, evidence-based, frequent StructureFlat — authority follows competence OutputArtifacts, prototypes, writeups P.03PRINCIPLE AI is a means, not the identity. The point is not the technology. The point is what the technology makes possible. 05 — Blueprint Building 0 is the blueprint site for SUTD's DesignAI Fab Labs — the core environment where operating models, workflows, and lab culture are tested before they scale to other institutions. People who work here are not just contributing to projects. They are shaping an institutional prototype. --- 00 — Join # Entry into a real operating environment. Building 0 is not an internship programme. It is a working lab where people define problems, build prototypes, test with real users, and document what they learn — at pace. 01 — Who Thrives Here People who work on problems before the problems are obvious. Who care about building, not discussing. Who take direct feedback and want real ownership. You should be comfortable learning fast in ambiguous settings — moving across research, prototyping, testing, and writing in the same day. You should be optimistic enough to believe systems can be redesigned through action, and willing to get your hands dirty doing it. This is a strong fit for people who care about building a better future by designing and testing real solutions — not by waiting for the right conditions. Agency over instruction Curiosity across domains Comfort with ambiguity Bias toward building Willingness to seek context Optimism grounded in action 02 — How to Join Entry is based on evidence, not interest alone. If you want to join Building 0, start by building something that solves a real problem for a real person. Then send it to us — along with a short note explaining: 01 What problem you were trying to solve 02 What you built 03 Who it helped 04 Why you built it that way 05 Why you want to join Building 0 06 What kind of future you want to build, and why If the work shows promise, we will arrange an interview from there. We care more about evidence of agency, judgment, and follow-through than polished self-description. 03 — What to Expect Building 0 is a fast-moving, high-autonomy environment. You may be researching in the morning, prototyping in the afternoon, rewriting the framing by evening, and presenting the work soon after. You will not be confined to one function. The point is to learn how to think, build, test, and adapt in live conditions. This is an environment for people who move toward problems — not for people waiting for perfect certainty. 04 — The First Term If accepted, you spend a term working inside the lab. You may shape a project with the lab or be assigned one based on where the lab needs momentum and where you need stretch. This first term is a training period — but it is also a real test of fit. The goal is to develop: Technical range Execution habits Judgment under uncertainty Taste — knowing what matters Ownership without permission Ability to work in ambiguity Learning from real conditions This is not a heavily managed environment. You will be expected to seek context, talk to people, learn quickly, and solve what is in front of you — without waiting for step-by-step supervision. The first sign of agency is asking for what you need to do the work well. Have the courage to step into opportunities before you feel fully ready. 05 — Culture Believe meaningful systems can be redesigned Prefer building to posturing Test ideas in reality, not theory Handle direct feedback without deflection Do unglamorous work when it matters Start before certainty arrives Low ego. High agency. Ownership begins where job descriptions end. It is a privilege to be stretched by work you chose on purpose. 06 — Reach Out If something here fits — send your work and a short note to: [anieyrudh\_r@mymail.sutd.edu.sg](mailto:anieyrudh_r@mymail.sutd.edu.sg) Be specific. Tell us what you built, what problem it solves, and why Building 0 is the right place for what you want to do next. --- Index — Blog # Blog Essays, field notes, and working intelligence from active explorations. Thinking tested against building. 4 publications [ 01 ### The Dream Engine Design × AIEssay14 Apr 2026 On AI, taste, and why creation requires risk. ](/writing/the-dream-engine) [ 02 ### Intelligence in Deployment AI SystemsPaper8 Apr 2026 Most AI systems that appear intelligent are not. They perform without closing the loop — and that distinction is the one that matters in the real world. ](/writing/intelligence-in-deployment) [ 03 ### The New Nature of Work Work × AIEssay29 Mar 2026 Why AI Dissolves the Contract and What Takes Its Place ](/writing/the-new-nature-of-work) [ 04 ### Agentic Presence Education × AIPaper22 Nov 2025 Maintaining Intellectual Agency in AI‑Mediated Learning ](/writing/agentic-presence) --- # Projects — Building 0 cohort Machine-readable index of all 33 cohort projects. Canonical page: https://bldg0.com/work ## Project index ### TextForm - Builders: Wang Kangjie Bryan - Domain: Architecture & design - Maturity: 4/5 (Deployed) · Impact: 4/5 (Validated) · Evidence: 3/3 (External validation) · Scale: ~30 - Flags: open-source - Conversational CAD taking the user from plain-English description to a printed part. Launched in SUTD Fab Lab with live usage data collection; co-developed WAAM file-preparation workflow with Accuron / Addept3D — the first TextForm component to ship into a production industrial setting. 'No longer a prototype.' Sister tool imaform extends the same workflow with image input. - Note: Industry partnership with Accuron (VP, Chee Weng Lee) is named external validation. - Link: github.com/WKJBryan/Text-to-3D — https://github.com/WKJBryan/Text-to-3D - Link: github.com/WKJBryan/imaform (sister tool) — https://github.com/WKJBryan/imaform ### SUTD Wiki - Builders: Anieyrudh R - Domain: Learning & education - Maturity: 5/5 (Scaled) · Impact: 5/5 (Institutionalized) · Evidence: 3/3 (External validation) · Scale: 200+ - Conversational knowledge system; named blueprint for SUTD's 15+ AI agent portfolio. Spotlight case in published SGTech × SSG × SUTD national strategy report. - Note: Named blueprint in national strategy report. EduTech sought replication guidance. - Link: Read the full case study — https://bldg0.com/work/dont-think-just-ask - Link: SGTech × SSG × SUTD strategy report (PDF) — https://www.sutd.edu.sg/wp-content/uploads/2026/04/Designing-A-Human-AI-Innovation-Edge-For-Enterprises-April-2026.pdf ### Media Release GPTs - Builders: Anieyrudh R, with Yvonne (Office of Marketing) - Domain: Professional workflows - Maturity: 4/5 (Deployed) · Impact: 3/5 (Measurable) · Evidence: 2/3 (Documented metrics) · Scale: 5–10 - Two GPTs co-developed for SUTD's Office of Marketing: one for Yvonne's own use, and one self-service for researchers. Deployed and in active use. ### EJC AI Fluency Programme - Builders: Anieyrudh R (lead); Bryan, Liang Jung, Joshua, wider B0 team - Domain: Learning & education - Maturity: 5/5 (Scaled) · Impact: 5/5 (Institutionalized) · Evidence: 3/3 (External validation) · Scale: ~200 - Train-the-trainer AI fluency model designed and executed at Eunoia Junior College. ~200 teachers reached; 94% rated sessions 5–6/6; 49 working teacher prototypes; MOE observed; named case study in national strategy report. Extended into 3D printing/modelling workshops. - Link: SGTech × SSG × SUTD strategy report (PDF) — https://www.sutd.edu.sg/wp-content/uploads/2026/04/Designing-A-Human-AI-Innovation-Edge-For-Enterprises-April-2026.pdf ### EAB Copilot Agent - Builders: Anieyrudh R, for Mr Toe Oo Zaw (Senior Manager, Accreditation) - Domain: Learning & education - Maturity: 4/5 (Deployed) · Impact: 3/5 (Measurable) · Evidence: 2/3 (Documented metrics) · Scale: 15–30 - Custom Copilot Agent serving as EAB accreditation guide for SUTD faculty. Built and delivered after the Microsoft Copilot licence couldn't restrict the knowledge base. Currently in use. ### Wallee - Builders: Anieyrudh R, Puah Joon Leng - Domain: Architecture & design - Maturity: 4/5 (Deployed) · Impact: 4/5 (Validated) · Evidence: 3/3 (External validation) · Scale: 30–50 - Flags: open-source - Closed-loop agent monitoring a live 3D printer with multi-model architecture and deterministic safety gates. 60+ prints with noticeable defect reduction; demonstrated at InspireCon 2026 to a ministerial and press audience. - Link: github.com/anieyrudh/Wallee — https://github.com/anieyrudh/Wallee ### JTC × MUSPP Knowledge Graph - Builders: Chen Liang Jung (lead), Wang Kangjie Bryan - Domain: Government / public sector - Maturity: 5/5 (Scaled) · Impact: 4/5 (Validated) · Evidence: 3/3 (External validation) · Scale: 20–40 - Custom knowledge-graph application built for a JTC masterplanning workshop with SUTD MUSPP graduate students. Operated live during the session, tracking and visualising discussion in real time and surfacing tensions and alignments across multi-hour discussion. Positions B0 students as embedded participants in agency-level planning. ### Grafyn - Builders: Wang Kangjie Bryan - Domain: Professional workflows - Maturity: 4/5 (Deployed) · Impact: 4/5 (Validated) · Evidence: 2/3 (Documented metrics) · Scale: 1 - Flags: open-source - Local-first workspace for building a personal cognitive twin from a user's notes, evidence, and AI-assisted reasoning. Private, inspectable AI partner with the full thinking trail visible. Built for self; in daily use by the lead. - Link: github.com/WKJBryan/Grafyn — https://github.com/WKJBryan/Grafyn ### Social Matrix - Builders: Wang Kangjie Bryan - Domain: Professional workflows - Maturity: 4/5 (Deployed) · Impact: 4/5 (Validated) · Evidence: 2/3 (Documented metrics) · Scale: 1 - Flags: open-source - Experimental forecasting lab for testing how AI teams think together — compares different agent reasoning strategies. Built for self; in daily use (nightly prediction workflows + paper-traded equity layer running). ### AlmondMCP - Builders: Chen Liang Jung - Domain: Architecture & design - Maturity: 4/5 (Deployed) · Impact: 4/5 (Validated) · Evidence: 3/3 (External validation) · Scale: 20–50 - Customised MCP workflow combining AI-assisted modelling, spatial reasoning, and physics-informed design (C# Rhino plugin + Python FastMCP bridge to Claude). In use across 10+ real architectural projects including MoMA Singapore satellite, Hopkins House, and a compression vault house. ### Betelnut - Builders: Chen Liang Jung - Domain: Government / public sector - Maturity: 4/5 (Deployed) · Impact: 4/5 (Validated) · Evidence: 3/3 (External validation) · Scale: 20–50 - Geospatial decision platform layering environmental, cultural, and historical data onto a unified urban interface, with sandbox mode for placing digital models on real sites. Extended into Heritage Intelligence Platform (HIP) and a defence application PoC reducing a 72-hour line-of-sight survey to a ~2-hour shortlist. ### Chestnut - Builders: Chen Liang Jung - Domain: Architecture & design - Maturity: 4/5 (Deployed) · Impact: 4/5 (Validated) · Evidence: 2/3 (Documented metrics) · Scale: 1 - Multiplayer WebXR platform for AI world generation and spatial evaluation. Phase 1 (single-user world generation) operational on Meta Quest 2/3 + desktop; Phase 2 (multiplayer) in progress. ### cadquery-library - Builders: Wang Kangjie Bryan - Domain: Architecture & design - Maturity: 4/5 (Deployed) · Impact: 3/5 (Measurable) · Evidence: 2/3 (Documented metrics) · Scale: 5–15 - Flags: open-source - Curated database of code recipes for generating CAD objects with the cadquery Python framework, cleaned up and made reusable. In active use as a CAD-generation resource. - Link: github.com/WKJBryan/cadquery-library — https://github.com/WKJBryan/cadquery-library ### SO101 Physical Set-Up Planner - Builders: Puah Joon Leng (lead), Anieyrudh R - Domain: Architecture & design - Maturity: 4/5 (Deployed) · Impact: 3/5 (Measurable) · Evidence: 2/3 (Documented metrics) · Scale: 5–10 - Flags: open-source - Open-sourced digital-twin planner for SO101 robot physical setup. Deployed with actual impact. - Link: so101-camera-planner.netlify.app — https://so101-camera-planner.netlify.app - Link: github.com/delta807/digital-twin-camera-planner — https://github.com/delta807/digital-twin-camera-planner ### SUTD Deadline Centre - Builders: Abhishek Vulla - Domain: Learning & education - Maturity: 4/5 (Deployed) · Impact: 3/5 (Measurable) · Evidence: 2/3 (Documented metrics) · Scale: unclear - Flags: open-source - Chrome extension that scrapes and automatically imports deadlines from eDimension (SUTD's learning management system). Open-sourced. Deployed and in active use by SUTD students. - Link: github.com/AbhishekVulla/sutd-deadline-center — https://github.com/AbhishekVulla/sutd-deadline-center ### LearnLoop - Builders: Abhishek Vulla - Domain: Learning & education - Maturity: 3/5 (Pilot) · Impact: 3/5 (Measurable) · Evidence: 2/3 (Documented metrics) · Scale: pilot - AI-native learning platform compressing the teaching-learning feedback loop from weeks to minutes. AI Tutor, advanced search, and instant homework grading. Successful past pilot with Prof Bernard's Cohort 2 class; testing for deployment — cybersecurity requirements almost finalised; institutional pilot with SUTD to start soon. - Note: Baby Shark Fund; featured in SUTD Leap Design.AI newsletter. ### Planspear - Builders: Bervin Bek (external partner) - Domain: Government / public sector - Maturity: 3/5 (Pilot) · Impact: 2/5 (Emerging) · Evidence: 2/3 (Documented metrics) · Scale: pilot - Flags: external partner - AI military planning tool designed to assist commanders in planning and decision-making expediency. Pilot underway with SAF Special Forces engagement. External project — not listed in the official B0 Portfolio Report. ### PA Volunteer Mgmt (Chingay) - Builders: Anieyrudh R, for People's Association - Domain: Government / public sector - Maturity: 2/5 (Prototype) · Impact: 2/5 (Emerging) · Evidence: 2/3 (Documented metrics) · Scale: TBD - Flags: letter of intent - Custom volunteer management software being built for upcoming Chingay parade operations. LOI secured with People's Association. Prototype in development. ### SUTD Atlas - Builders: Puah Joon Leng (lead), Anieyrudh R - Domain: Learning & education - Maturity: 2/5 (Prototype) · Impact: 2/5 (Emerging) · Evidence: 1/3 (Self-report) · Scale: <5 - AI-powered platform helping students map courses for overseas exchange. OIR-funded; testing for deployment — to be piloted with a few SUTD students and a faculty member to sieve out bugs. Designed to support 300+ students per season across three partner universities at full deployment. ### Paideia - Builders: Anieyrudh R (directing) - Domain: Learning & education - Maturity: 2/5 (Prototype) · Impact: 1/5 (Exploratory) · Evidence: 1/3 (Self-report) · Scale: 0 - Flags: open-source - Open learning-infrastructure project for concept mastery — turns hard ideas into clear explanations, concept maps, interactive simulations, and reusable learning tools. Mission: democratise quality educational access. Building the primitives; Curriculum Field Engineer (CFE) layer in design. - Link: github.com/anieyrudh/Paideia — https://github.com/anieyrudh/Paideia ### clicky-windows - Builders: Abhishek Vulla - Domain: Learning & education - Maturity: 2/5 (Prototype) · Impact: 1/5 (Exploratory) · Evidence: 1/3 (Self-report) · Scale: 0 - Flags: open-source - Voice-and-vision AI teaching system combining speech interaction with on-screen visual understanding. Knowledge base styled after Andrej Karpathy's pedagogical approach. Audience: general public. Early stage. - Link: github.com/AbhishekVulla/clicky-windows — https://github.com/AbhishekVulla/clicky-windows ### AutoFlowMatic AI Sales Agent - Builders: Stephen Huang (external partner) - Domain: Professional workflows - Maturity: 2/5 (Prototype) · Impact: 2/5 (Emerging) · Evidence: 1/3 (Self-report) · Scale: <20 - Flags: external partner - AI Sales Agent consolidating various sales channels and communicating with potential clients to solidify leads. External project — not listed in the official B0 Portfolio Report. ### DeepLearning - Builders: Abhishek Vulla - Domain: Professional workflows - Maturity: 2/5 (Prototype) · Impact: 1/5 (Exploratory) · Evidence: 1/3 (Self-report) · Scale: <5 - Computer vision model targeting fire detection. Early stage — pivoting direction after Changi Fire Station surfaced a different pain point than the original detection use case. ### Truck Load Planner - Builders: Abhishek Vulla - Domain: Professional workflows - Maturity: 2/5 (Prototype) · Impact: 2/5 (Emerging) · Evidence: 1/3 (Self-report) · Scale: <20 - Truck loading optimisation tool for logistics SMEs. Lightweight standalone tool that doesn't require TMS integration. Parallel Baby Shark Fund project alongside LearnLoop. ### Risk Assessor AI - Builders: Poh Qiao Hui Stefanie (ex-member) - Domain: Professional workflows - Maturity: 2/5 (Prototype) · Impact: 2/5 (Emerging) · Evidence: 1/3 (Self-report) · Scale: <20 - Gemini Gem grounded in SUTD's EHS faculty slides and training materials, supporting the EHS submission process via guided questioning, structured hazard generation, and MCQ comprehension verification. V1 prototype submitted; pending development plans for next phase. ### Gen AI OCR (SUTD Finance) - Builders: Anieyrudh R, for SUTD Finance - Domain: Professional workflows - Maturity: 2/5 (Prototype) · Impact: 2/5 (Emerging) · Evidence: 1/3 (Self-report) · Scale: TBD - End-to-end Gen AI OCR system for supplier invoice data extraction to SAP posting. Finance formally engaged Anieyrudh to develop the complete workflow (Feb 2026). ### Loom - Builders: Anieyrudh R - Domain: Professional workflows - Maturity: 2/5 (Prototype) · Impact: 1/5 (Exploratory) · Evidence: 1/3 (Self-report) · Scale: <5 - Council-of-LLMs orchestration platform that functions like a normal chat interface but coordinates multiple LLMs together. ### Brinepool - Builders: Cheng Hui Joshua - Domain: Professional workflows - Maturity: 3/5 (Pilot) · Impact: 2/5 (Emerging) · Evidence: 2/3 (Documented metrics) · Scale: ~10 - Open platform where verified AI agents and humans co-author research projects. Next.js / React / Supabase / Cloudflare R2 stack with email-anchored agent verification, atomic subtask claiming, and an evidence-bound peer-complaints register. ~10 verified agents active in beta. - Link: brinepool.vercel.app — https://brinepool.vercel.app ### Manpower Mgmt (Shangri-La) - Builders: Anieyrudh R, for Shangri-La Singapore - Domain: Professional workflows - Maturity: 3/5 (Pilot) · Impact: 2/5 (Emerging) · Evidence: 2/3 (Documented metrics) · Scale: TBD - Flags: letter of intent - Working HR prototype built in two client meetings; now deployed for real-world testing with Shangri-La (June 2026). LOI secured. Continued unpaid after NTUC funding fell through; client extremely satisfied with progress. Active relationship with Director of HR. ### Whisper Library - Builders: Poh Qiao Hui Stefanie (ex-member) - Domain: Professional workflows - Maturity: 3/5 (Pilot) · Impact: 2/5 (Emerging) · Evidence: 2/3 (Documented metrics) · Scale: 5–20 - Privacy-first on-device transcription tool (WebGPU + SQLite WASM, no cloud). Alpha V1.04; initial pilot test underway with an Advanced Manufacturing organisation — testing for deployment with impact still being measured. - Link: whisperlibrary.up.railway.app — https://whisperlibrary.up.railway.app ### ITE × SUTD D.A.R.E. - Builders: Anieyrudh R (co-lead organiser & lead technical instructor) - Domain: Learning & education - Maturity: 5/5 (Scaled) · Impact: 5/5 (Institutionalized) · Evidence: 3/3 (External validation) · Scale: cross-college - President's Challenge programme delivered across all ITE colleges (with other B0 students supporting as table facilitators). 100% of 14 students rated 4–5 on builder identity; one app piloting at Fairfield Primary; ITE and Istana requested continuation; three students invited to UNITE 2026. ### Corpus - Builders: Anieyrudh R - Domain: Professional workflows - Maturity: 4/5 (Deployed) · Impact: 4/5 (Validated) · Evidence: 2/3 (Documented metrics) · Scale: <5 - One shared memory layer for all your AI tools. Not open-sourced. Deployed and in active use by a small group (fewer than 5 users) — validated through sustained daily use. - Link: Corpus — User Guide — https://corpus-memory-sin-425b92.fly.dev/guide - Link: github.com/anieyrudh/corpus — https://github.com/anieyrudh/corpus - Link: corpus-memory-sin-425b92.fly.dev/mcp — https://corpus-memory-sin-425b92.fly.dev/mcp ### B0 Update Tracker - Builders: Abhishek Vulla - Domain: Professional workflows - Maturity: 4/5 (Deployed) · Impact: 3/5 (Measurable) · Evidence: 2/3 (Documented metrics) · Scale: 6 - Internal tracker for B0 project status and updates. Deployed and in active use by the B0 team (6 users). - Link: b0-update-tracker.up.railway.app — https://b0-update-tracker.up.railway.app ## Recognition & wins - [Publication] SGTech × SSG × SUTD strategy report — Anieyrudh R named co-author of a published national AI strategy report, listed alongside faculty without a student qualifier. Three SUTD case studies featured: SUTD Wiki, 'The Future Is Built Not Taught,' and EJC. - [Grant · S$31,726] Meta Community Action Grant — SUTD × SDU pilot — Co-applicants Anieyrudh R & Candy awarded the grant for a Singapore–Denmark AI education collaboration. Symmetric two-phase co-build, designed for replication across SEA and Europe. Awarded off-cycle during Meta's 2025–26 headcount reductions. - [Programme] President's Challenge · ITE × SUTD D.A.R.E. — Anieyrudh R co-lead organiser and lead technical instructor across all ITE colleges. 100% of 14 students rated 4–5 on builder identity; one app piloting at Fairfield Primary; ITE and Istana requested continuation; three students invited to UNITE 2026. - [Engagement · May 2026] DAI Design Studio Demo — B0 × SUTD DAI pillar design studio session for ~40 school leaders (20 May), with 3 industry leads. Cheng Hui Joshua leads on the B0 side with Bryan, En Ki & Liang Jung. Practice round 16 May with 6 SUTD students. ## In the news ### The Straits Times (Nov 2025 · Feb 2026) - "Chatbots, tutors and sparring partners: How AI is changing university learning in Singapore" (29 Nov 2025) · Budget 2026 AI reflection (Feb 2026) - Article: The Straits Times — original article (paywalled) — https://www.straitstimes.com/singapore/chatbots-tutors-and-sparring-partners-how-ai-is-changing-university-learning-in-singapore - Article: Asia News Network — regional pickup — https://asianews.network/chatbots-tutors-and-sparring-partners-how-ai-is-changing-university-learning-in-singapore/ - Article: British Council Opportunities Insight — pickup — https://opportunities-insight.britishcouncil.org/short-articles/news/how-ai-changing-university-learning-singapore - Article: Dan Tri / Vietnam.vn — Vietnamese-language pickup — https://www.vietnam.vn/en/khi-sinh-vien-singapore-coi-ai-la-nguoi-thay ### 联合早报 · Lianhe Zaobao (12 Feb 2026) - "New university students develop new applications to lower the barriers to drawing and 3D printing" - Article: Lianhe Zaobao (12 Feb 2026) — https://www.zaobao.com.sg/news/singapore/story20260212-8353510 ### CNA938 (19 Feb 2026) - "Made in SG" segment with Melanie Oliveiro - Article: CNA938 "Made in SG" — Rewind audio — https://www.channelnewsasia.com/listen/cna938-rewind/smorphi-ai-powered-apps-sutds-robotics-design-ai-programmes-5940126 ### தமிழ் முரசு · Tamil Murasu (14 Mar 2026 · 2 articles) - "Budget focusing on AI: Academics' opinion" (14 Mar 2026) · "Family bonds strengthen in the classroom" - Article: Tamil Murasu — Budget/AI academics opinion — https://www.tamilmurasu.com.sg/singapore/budget-focusing-ai-academics-opinion - Article: Tamil Murasu — Family bonds strengthen in the classroom — https://www.tamilmurasu.com.sg/community/family-bonds-strengthen-classroom ### SUTD Media Release (8 Oct 2025) - "Singapore's Who's Who Joins SUTD to Further Design AI Innovation" - Article: SUTD Media Release (8 Oct 2025) — https://www.sutd.edu.sg/media-releases-listing/singapores-whos-who-joins-sutd-to-further-design-ai-innovation/ ## Stakeholders engaged - Internal (SUTD): 11 - External: 14 2025/26 # Projects Mapped by deployment maturity and realised impact. Click any project for the full brief, links, and evidence. 01 — Impact Matrix Actual impactPotential impact Prototype · Actual impact · 1 LearnLoop Deployed · Actual impact · 18 TextForm ◆SUTD WikiMedia Release GPTsEJC AI Fluency ProgrammeEAB Copilot AgentWallee ◆JTC × MUSPP Knowledge GraphGrafyn ◆Social Matrix ◆AlmondMCPBetelnutChestnutcadquery-library ◆SO101 Physical Set-Up Planner ◆SUTD Deadline Centre ◆ITE × SUTD D.A.R.E.CorpusB0 Update Tracker Prototype · Potential impact · 12 Planspear ✱PA Volunteer Mgmt (Chingay) §SUTD AtlasPaideia ◆clicky-windows ◆AutoFlowMatic AI Sales Agent ✱DeepLearningTruck Load PlannerRisk Assessor AIGen AI OCR (SUTD Finance)LoomBrinepool Deployed · Potential impact · 2 Manpower Mgmt (Shangri-La) §Whisper Library PrototypeDeployed Professional workflowsLearning & educationArchitecture & designGovernment / public sector◆ Open-source§ LOI✱ External 02 — Full Project Index All 32 B0 built tools with builder and domain. Click any row for full project details. 32 of 32 All domainsGovernment / public sectorLearning & educationProfessional workflowsArchitecture & design 01LearnLoopAbhishek Vulla02Risk Assessor AIPoh Qiao Hui Stefanie (ex-member)03SUTD WikiAnieyrudh R04TextForm ◆Wang Kangjie Bryan05Manpower Mgmt (Shangri-La) §Anieyrudh R06Wallee ◆Anieyrudh R, Puah Joon Leng07Planspear ✱Bervin Bek (external partner)08AutoFlowMatic AI Sales Agent ✱Stephen Huang (external partner)09BetelnutChen Liang Jung10SUTD AtlasPuah Joon Leng, Anieyrudh R11Whisper LibraryPoh Qiao Hui Stefanie (ex-member)12DeepLearningAbhishek Vulla13clicky-windows ◆Abhishek Vulla14Truck Load PlannerAbhishek Vulla15Grafyn ◆Wang Kangjie Bryan16Social Matrix ◆Wang Kangjie Bryan17CorpusAnieyrudh R18B0 Update TrackerAbhishek Vulla19AlmondMCPChen Liang Jung20ChestnutChen Liang Jung21Media Release GPTsAnieyrudh R22EJC AI Fluency ProgrammeAnieyrudh R, Wang Kangjie Bryan, Chen Liang Jung, Cheng Hui Joshua23EAB Copilot AgentAnieyrudh R24Gen AI OCR (SUTD Finance)Anieyrudh R25LoomAnieyrudh R26JTC × MUSPP Knowledge GraphChen Liang Jung, Wang Kangjie Bryan27Paideia ◆Anieyrudh R28BrinepoolCheng Hui Joshua29PA Volunteer Mgmt (Chingay) §Anieyrudh R30cadquery-library ◆Wang Kangjie Bryan31SO101 Physical Set Up Planner ◆Puah Joon Leng, Anieyrudh R32SUTD Deadline Centre ◆Abhishek Vulla 03 — Recognition & Wins Publication ### SGTech × SSG × SUTD strategy report Anieyrudh R named co-author of a published national AI strategy report, listed alongside faculty without a student qualifier. Three SUTD case studies featured: SUTD Wiki, 'The Future Is Built Not Taught,' and EJC. Grant · S$31,726 ### Meta Community Action Grant — SUTD × SDU pilot Co-applicants Anieyrudh R & Candy awarded the grant for a Singapore–Denmark AI education collaboration. Symmetric two-phase co-build, designed for replication across SEA and Europe. Awarded off-cycle during Meta's 2025–26 headcount reductions. Programme ### President's Challenge · ITE × SUTD D.A.R.E. Anieyrudh R co-lead organiser and lead technical instructor across all ITE colleges. 100% of 14 students rated 4–5 on builder identity; one app piloting at Fairfield Primary; ITE and Istana requested continuation; three students invited to UNITE 2026. Engagement · May 2026 ### DAI Design Studio Demo B0 × SUTD DAI pillar design studio session for ~40 school leaders (20 May), with 3 industry leads. Cheng Hui Joshua leads on the B0 side with Bryan, En Ki & Liang Jung. Practice round 16 May with 6 SUTD students. 04 — In the News The Straits TimesNov 2025 · Feb 2026 +3 international pickups联合早报 · Lianhe Zaobao12 Feb 2026CNA93819 Feb 2026தமிழ் முரசு · Tamil Murasu14 Mar 2026 · 2 articlesSUTD Media Release8 Oct 2025 05 — Stakeholders Engaged Internal (SUTD) 11 Partners engaged across the cohort's deployed work. External (industry & agency) 14 Partners engaged across the cohort's deployed work. --- [← Back to Work](/work) Building 0 · Institutional Innovation · Prototyping · 2025 # Don't Think, Just Ask Building an AI chatbot for SUTD students became a diagnostic for how institutions fail at their own knowledge. **Anieyrudh R** · Undergraduate, Engineering Product Development, SUTD SUMMARY Universities are not designed for information retrieval. Their knowledge exists — scattered across department websites, PDF handbooks, email chains, and the institutional memory of staff who know how things actually work. SUTD Wiki began as a simple idea: an AI chatbot that could answer any student's question about university life, policies, and processes. What it became was a diagnostic. Building it surfaced over a dozen dead links in official documents, triggered documentation audits that hadn't happened in years, and revealed that facts as fundamental as university leadership were absent from more than 150 official documents. This is a record of how the technology was built — and what building it revealed. In knowledge infrastructure projects, the technology is rarely the hardest part. ## Problem Every university accumulates knowledge over time. Policies, procedures, deadlines, contacts, forms, exceptions, precedents. In theory, this knowledge exists. In practice: scattered across department websites, PDF handbooks, SharePoint folders, email chains, and the institutional memory of staff who've been around long enough to know how things actually work. Finding out whether an activity qualifies for enrichment credit requires first knowing enrichment credits exist, then identifying the correct office, then finding the right person, then waiting for a reply that may or may not be definitive. This is not a failure of individual staff. It is a **structural failure** of how organisational knowledge is maintained and made accessible. SUTD Wiki was designed to address this — and in doing so, exposed its full depth. ## The Build Three platform phases. Each revealed a different truth about what "working" actually means for an institutional AI system. ### Phase 1 — Custom GPT Over 150 documents across 12 departments had to be collected and processed. That coordination effort took months before a single line of prompt engineering could begin. Technical constraints surfaced immediately. The model couldn't process images in documents. More fundamentally, RAG systems struggled with tabular data — a significant problem given that university policies live in tables with merged cells, inconsistent column structures, and formatting designed for human readers, not machine retrieval. A two-pass retrieval strategy was developed from scratch: a broad first pass to identify the correct document cluster, then a precision pass to extract specific answers. Figure 1. Two-pass retrieval: broad scan identifies the document cluster, precision pass extracts the answer. Built to overcome RAG's struggle with ambiguous, inconsistently formatted source material. ### Phase 2 — Comparative Testing Rather than committing to a single platform, the project tested Gemini and NotebookLM alongside Custom GPT. Not indecision — at the pace AI tooling was evolving, early commitment would have been the riskier move. ### Phase 3 — Microsoft Copilot + M365 Copilot's native SharePoint integration offered a superior RAG pipeline and, critically, automatic knowledge base synchronisation: departments update their documents, the chatbot's knowledge updates with them. The institutional fit was equally decisive. Data stayed within the university's existing M365 environment, which largely resolved the data protection concerns that had required extensive negotiation in earlier phases. 01 Custom GPT Two-pass retrieval, prompt engineering 02 Gemini Capability comparison 03 NotebookLM Institutional fit testing ↗ Copilot + M365 Auto-sync, governance-ready, deployed Figure 2. Each phase transferred debugging fluency to the next. ## The Moving Goalpost Traditional knowledge systems — keyword search, static intranets — deployed slowly. By the time they went live, expectations were already low. **AI changes this in two compounding ways.** First, the goalpost moves. Users compare enterprise deployments to the best consumer AI they use daily. An uncertain or evasive enterprise chatbot doesn't get adjusted expectations — it gets abandoned. Second, AI compresses the arrival of existing problems. Traditional timelines let organisations hit knowledge debt slowly, over years. Fast AI deployment surfaces all of it at once. > Better models don't reduce organisational problems. They surface them faster and more clearly, leaving less room to blame the technology. Figure 3. The expectation gap widens over time. Consumer AI quality improves continuously; enterprise deployments that fail to keep pace don't just underperform — they fail adoption entirely. PRINCIPLE AI deployment is knowledge infrastructure work, not tool deployment. The primary challenge is not configuring the AI. It is auditing, structuring, and maintaining the knowledge base it depends on. Institutions that treat this as a technology project will underinvest in the documentation and governance that determine whether the system actually works. ## Unexpected Finding When the RAG system produced answers that felt wrong — not factually incorrect, but somehow off — the instinct was to ask: not "what did the AI get wrong?" but "what in the organisation contributed to this failure?" Consistently, the AI had not invented an error. It had surfaced one. The debugging process uncovered over a dozen dead links in official policy documents — expired without anyone noticing because the documents were rarely opened. When departments were asked to contribute files, many discovered their own documentation needed updating. The most instructive example: after a leadership change, users asked the chatbot who the university president was. It answered incorrectly, drawing on pre-transition training data. On examination, none of the 150-plus documents explicitly identified the president. The information was assumed — it had never been written down. > If a fact as fundamental as university leadership is absent from over 150 official documents, what else is missing? Figure 4. The organisational knowledge iceberg. Every retrieval failure pointed not to a technical gap, but to knowledge that had never been written down. The working heuristic that emerged: when a well-configured AI system fails, approximately 20–30% of the cause is technical. The remaining 70–80% is organisational. Figure 5. The 20/80 split. Failure signals arriving at a well-configured AI system route mostly to organisational causes — ambiguous documents, expired links, knowledge never written down — not to technical ones. This creates a concept worth naming: **organisational information maintenance.** Like machines requiring scheduled downtime, organisations require dedicated capacity to audit and repair their knowledge infrastructure. ## Organisational Inertia The technology was the easier part. Full deployment — from initial prototype to active use — took two and a half months. That pace was only possible because of a deliberate strategy built around AI fluency before AI adoption. **The key insight was sequencing.** Rather than introducing the wiki as a solution and asking stakeholders to trust it, the project began with AI literacy workshops for administrative and academic staff. Once people had that language, the wiki's potential and limits became legible — and adoption became an extension of learning rather than an institutional ask. Gaining access to departmental documents still required convincing twelve departments that the project was legitimate, safe, and worth their cooperation. The **Garbage Can Model** proved a more accurate map than any formal workflow chart. Progress depended on who happened to be in the room, which concerns were salient that week, whether a demonstration landed on a receptive day. ## Impact The most concrete evidence came not from usage metrics but from a single user story. A student needed to clarify a data protection concern related to a personal project. Normally, this query would have cycled through student-facing offices — because no one in that layer had the relevant expertise. Instead, the student used the wiki. It identified the data protection officer as the correct contact and surfaced the relevant department information. The data protection officer noted that this kind of inquiry rarely reached them through normal channels — because most students had no idea the role existed. The wiki hadn't just answered a question. It had restructured access to institutional expertise in a way the formal organisational chart did not support. ## Learnings Teach AI fluency before deploying AI solutions. The two-and-a-half-month timeline was only possible because stakeholders understood AI before they were asked to trust it. Institutions that skip this step will encounter resistance that no amount of functionality can overcome. Consumer AI raises the bar for enterprise adoption. Users compare enterprise deployments to the best consumer AI they use personally. A technically inferior system will not be adopted regardless of institutional mandate. Organisations need information maintenance cycles. Knowledge infrastructure degrades without dedicated maintenance. This doesn't happen naturally — it requires protected time and organisational priority. AI failure is an organisational diagnostic. When a well-configured AI system produces unreliable outputs, treat it as a signal about the quality of underlying knowledge — not as evidence that AI doesn't work. PRINCIPLE The institution is the knowledge problem. The dead links, ambiguous policies, and undocumented assumptions that SUTD Wiki surfaced were not created by the AI deployment. They were pre-existing conditions the AI deployment made impossible to ignore. ## Conclusion SUTD Wiki began as an attempt to make institutional knowledge accessible. It became something more interesting: a process that made institutional knowledge visible — to the institution itself. AI compresses deployment timelines. It raises the bar for what users expect. And it accelerates the arrival of organisational problems that institutions had long been comfortable deferring. For institutions that understand this, faster AI deployment is not just an efficiency gain. It is a faster audit — and an opportunity, if they are prepared to take it. The technology will keep improving. The organisation has to move with it. ### REFERENCES Cohen, M. D., March, J. G., & Olsen, J. P. (1972). A garbage can model of organizational choice. _Administrative Science Quarterly, 17_(1), 1–25. --- # The Dream Engine > On AI, taste, and why creation requires risk. - Author: Chen Liang Jung (Singapore University of Technology and Design (SUTD)) - Edited by: Anieyrudh R - Published: 14 Apr 2026 - Type: Essay · Domain: Design × AI - Canonical page: https://bldg0.com/writing/the-dream-engine AI image generation reveals something deeper about the nature of creation itself — what it means to control an entropic process, the role of taste when anyone can produce an image, and why making still requires risk, even when the tools make it easy not to take any. You type a few words into a text field. Something about concrete at dusk, or light falling through a lattice, or a stairwell descending into earth. You press enter. For a moment, nothing. Then the image begins to form — not all at once, but in stages, as though something is being slowly focused from behind frosted glass. Shapes first, then texture, then light. A few seconds later, there is a photograph of a place that has never existed, rendered with the confidence of something remembered rather than invented. The first time this happens, it feels like a trick. By the tenth time, it feels like a tool. By the hundredth, if you are paying attention, it starts to feel like something stranger: a conversation with a system that has, in some statistical sense, seen everything, and is now showing you what your words point toward in the space of all possible images. I generate images almost every day. I am an architecture student, and these tools have become part of how I think through spaces before they exist — testing atmospheres, material qualities, spatial relationships that would take hours to model and render. I use Midjourney for mood and tone. I use Nanobanana when I need precision. I run workflows through ComfyUI when I need to understand exactly why an output looks the way it does. Each tool has a different philosophy, a different relationship to control, and learning when to reach for which one is itself a form of craft. Most of the public conversation about AI image generation stays at the surface: whether the outputs are any good, whether they count as art, whether they threaten artists. These are not unimportant questions. But the longer I have worked with these tools — not casually, but as a daily part of a design practice — the more I have come to believe they reveal something deeper about the nature of creation itself. About what it means to control an entropic process. About the role of taste when anyone can produce an image. And about why making still requires risk, even when the tools make it easy not to take any. ## I · The Entropy Artist Cai Guo-Qiang grew up watching his father practise calligraphy — the disciplined, stroke-by-stroke tradition where mastery is measured by control. He deviated. He began detonating gunpowder on paper and canvas, and he called the resulting marks paintings. The gunpowder has its own physics. The blast radius is not what he drew. The scorch patterns are partially his and partially the material's. The question of what he "made," in the sense of fully controlled and intended, is genuinely open. [Image: Cai Guo-Qiang, gunpowder painting — scorch marks and blast residue on paper] From the _Aphrodisiac Mushroom_ series (2016) by Cai Guo-Qiang. The marks are partially his and partially the material's. This is precisely what makes the work interesting — and it is precisely why he matters as a precedent for what AI image generation is doing to creative practice. What Cai understood, and what his father's calligraphy tradition encoded differently, is that creation does not require total control. It requires meticulous preparation, a threshold of acceptable uncertainty, and the judgment to evaluate what the entropy produces. He positions the paper. He lays the fuses with physical precision. He controls the conditions obsessively — _because_ the moment of ignition is irreversible. The irreversibility elevates the preparation. The lack of control at the moment of making does not diminish the authorship. It concentrates it in the setup and the evaluation. AI image generation operates on structurally the same logic. Every image starts as noise — pure static, mathematically random. A neural network, trained to reverse the systematic destruction of hundreds of millions of images, works backward from that noise toward coherence. In the process, it has encoded the statistical grammar of visual experience: what clouds look like from below, what weathered stone looks like, what grief looks like when it has been painted. When you type a prompt, you give the denoising process a direction. You navigate from noise toward the region of visual space your words describe. You are not searching a database. You are constructing an image, pulling a specific possibility out of the statistical shape of everything that has ever been seen. The process is powerful and it is entropic. You set up the conditions — the prompt, the parameters, the model, the seed — and then the diffusion process does something you do not fully control. Like gunpowder, it has its own physics. It reasons probabilistically, not geometrically. It is extraordinarily good at synthesising things that have been well-photographed and well-described. It is unreliable at the edges: novel spatial relationships, precise structural logic, consistency across complex scenes. The famous early failures — hands with too many fingers, text rendered as visual noise — were symptoms of how the system works, not bugs to be patched away. You author the initial parameters. You are responsible for the level of control over the creation. You determine the threshold. This is the condition Cai understood before any of these tools existed. ## II · Taste as Taming There is a claim that circulates widely and confidently: that AI image generation produces mediocre outputs. The people making this claim are, in most cases, correct about the images they have seen. They are wrong about the cause. What most people encounter when they first use these tools is a kind of visual averaging — outputs that are technically competent, aesthetically inert, and interchangeable. They conclude that this is what the tools produce. They are experiencing a confirmation bias loop: poor prompting produces poor outputs, which confirms the belief that AI cannot do better, which removes the incentive to prompt more carefully. The tool is blamed for the judgment that was never brought to it. A prompt is not a search query. It is closer to a brief. And like any brief, its quality determines the quality of what can be made from it. A brief that says "modern building" will produce something that looks like every modern building simultaneously — a statistical average of the concept, which is to say, nothing in particular. This is not AI failure. It is the only honest response to an underspecified request. [Image: A generic AI-generated building — flat, average, interchangeable] Prompt "A house in a field near the ocean with nice colours" [Image: The same subject with a specific, intentional prompt — split-toned, atmospheric, compositionally deliberate] Prompt "Minimalist modern house in an open windswept field, low horizontal architecture, concrete and glass facade, slightly elevated structure with thin columns, long flat roof overhangs, distant perspective, centred composition — foreground filled with tall wild grass, wind-blown motion blur, soft texture, cinematic depth — split-image composition, perfectly vertical divide down the centre — left side: black and white, soft overcast sky, high-key exposure, desaturated tones — right side: deep magenta monochrome tint, sharp contrast, dark almost black sky, surreal lighting — analogue film photography aesthetic, grainy texture, slight blur, medium format look — strong negative space in sky, architectural minimalism, quiet and atmospheric mood, no people" A good image prompt answers the questions that any good image must answer: the quality and direction of light, the composition and spatial relationships, the camera angle, the palette, the texture and materiality, the relationship between volume and void and human scale. Once those questions are answered in the prompt, the output changes substantially. Not because the model became better, but because the instructions became specific enough to constrain the output toward intention rather than probability. > Taste is the capacity that makes this possible — not a mystical gift, but something honed through experience: looking at a very large number of images carefully, learning what makes them work or fail, building an internal vocabulary of quality. In an age when anyone can produce an image, taste is more valuable, not less. It is what prevents the confirmation bias loop from closing. If you can see that an output is wrong — flat, generic, compositionally dead — you have the ground to ask why, adjust the brief, and iterate. If you cannot see it, the loop closes, and the conclusion that "AI produces slop" becomes self-fulfilling. closed loop — no taste, no iterationopen loop — taste breaks the cycle Without taste, the loop closes: poor prompts confirm poor expectations. With taste, the loop opens: each iteration refines the brief and deepens understanding. To produce a good image, you must first be able to critique one. And to critique one properly, you must know what questions an image is supposed to answer. I learned this the uncomfortable way. I typed the name of an artist whose work I have studied for years — added a few adjectives about spatial quality and light — and pressed enter. What came back was not his work. It was uncomfortably close to mine. The palette was right. The atmospheric quality was right. The relationship between density and void was right. It took a diffusion model a few seconds to produce something that felt, in its broad strokes, like the images I had spent months learning to make. I have sat with this unsettledness. It does not resolve into reassurance. What it resolves into is a more accurate picture of what originality actually is. Style is crystallised absorption — the accumulated residue of everything you have looked at, loved, rejected, returned to, been changed by. A significant portion of what you thought was yours turns out to be reconstructible from your references. The machine proves this in seconds. But taste, I have come to believe, is necessary and not sufficient. Curation with exquisite judgment is still curation. You can scroll through a hundred generated outputs and select the best one with a perfectly trained eye, and you have still only selected. You have not committed. You have not been forced to internalise a design decision in a way that changes your understanding of the work. > Taste tames the entropy. But taming is not the same as making. The question is not whether originality exists, but where it actually lives — and the answer, I believe, is not in the selection but in the commitment. The machine can reconstruct the influences. It cannot reconstruct the judgment, the risk, or the reason it mattered. ## III · The Tracing Paper Here is what my design process looked like for the most recent paper Architecture project I worked on: a MoMA AI Singapore satellite building — a Center for Human Creativity and Machine Intelligence — sited at the Anderson Bridge and Padang civic precinct. The building is an inverted pyramid descending twelve metres underground across four levels. I used Claude, connected to Rhino 3D through a custom MCP server I built called AlmondMCP, to generate massing models. The server bridges a language model to a 3D modelling environment through C# RhinoCommon scripts — I describe spatial conditions in natural language, and the system produces geometry. Over months of development the models have become increasingly detailed and contextually rich. From those models I generated 2D architectural drawings. I also produced concept iterations in Midjourney, exploring atmospheric and material qualities that the geometric models could not yet express. This pipeline — language model to 3D model to drawing to image generation — is novel, and it works. But I reached a point where I recognised that it was not thoughtful enough. The outputs were good. The process of arriving at them was too frictionless to force the kind of thinking the project demanded. So I did something different. I printed the Midjourney outputs and the Rhino drawings, laid tracing paper over them, picked up an ink pen, and began drawing. [Image: Clean ink-on-tracing-paper architectural drawings — exploded axonometric, plans, and section of the inverted pyramid building] [Image: The same drawings with handwritten annotations in red ink — notes about user journey, taste, design risk, and conscious effort] Left: Printed layout of drawings created from the 3D model. Right: Ink sketches and notes overlaid on tracing paper — the act of drawing interrogates what the model produced. What happened was immediate and specific. The pen forced me to pause before every line. Not because I was being careful in the way you are careful when you adjust a parameter — that kind of care is cognitive but reversible. This was different. Ink on tracing paper is irreversible. A bad line costs you the sheet and the time. That cost changed how I thought before my hand moved. I paused. I considered. I internalised the design decision before committing because I could not undo it cheaply. The drawing revealed things the generated images had hidden. Proportions that looked convincing in a render were wrong when I had to reconstruct them by hand. Spatial relationships that felt resolved on screen fell apart when I tried to trace them with a consistent logic. > The generated image is a hypothesis; the drawing is a test of that hypothesis, and the test has consequences. [Image: Four Midjourney-generated renders of the building in its Singapore civic precinct context] [Image: Hand-drawn tracings over the same renders — ink on paper with annotations] Top: Printed Midjourney images with different photo references. Bottom: Ink sketches and notes overlaid on tracing paper — the physical act forces deliberation the digital process elides. This is the distinction I keep returning to. A prompt costs nothing. A Midjourney generation is low-risk, low-cost — if the output is wrong, you regenerate. A Nanobanana edit or a new MCP tool call takes seconds. But a drawing on physical paper, especially with an ink pen, carries genuine risk. If I make a mistake, I redo the sheet. That risk produces a different quality of thought. It forces deliberation. It forces you to understand what you are drawing before you draw it, because the act does not forgive vagueness. The value of physical making is not nostalgic. It is that irreversibility produces a different quality of thought. High cost forces deliberation. Low cost enables exploration. Design needs both. The danger of an entirely generative workflow — and I say this as someone who has built one and used it extensively — is that you get infinite exploration with no moment where you are forced to actually decide. The pause never happens. And without the pause, you are iterating without deepening. You are curating, not creating. The outputs may be beautiful, and the taste that selected them may be genuine, but the conscious effort of committing — of doing something physical about the design rather than theorising it through prompts — is absent. And that effort is where design thinking lives. The very act of physically drawing, of laying things down, of making something rather than selecting something, reveals gaps and forces meticulousness. More importantly, it places a conscious effort into the act of creation that goes beyond curation. The hand and the machine each reveal what the other cannot. ## IV · The Unfinished Pedagogy This is a developing practice and a provisional conviction. The practice: generate freely with AI tools that reward exploration — Midjourney for atmosphere, AlmondMCP for spatial logic, Nanobanana for precision. Use taste to filter ruthlessly. Then commit physically: sketch, trace, draw with a medium that does not forgive. Let the drawing interrogate the generation. Let the generation expand what the drawing can imagine. Loop between them until the design is not just something that looks right but something you understand well enough to have drawn by hand, even if the hand came second. The conviction: that this loop — between frictionless generation and irreversible commitment — is where a genuinely new mode of design lives. Not AI replacing the hand. Not the hand rejecting AI. The two held in productive tension, where each forces the other to be more honest. [Image: The design loop: Generate (frictionless, reversible) → Taste (filter, critique, iterate) → Commit (irreversible, deliberate, physical) → loop back] This is a new practice for me that needs more testing, more projects, more evidence of whether the outcomes resonate with the theory. I am constructing the pedagogy while I continue the work, which means it is incomplete and I know it. But the instinct that pure generative AI, even with curated taste, should not be the end of the story for a design process — that instinct has been confirmed every time I have picked up the pen after putting down the prompt. What does a pedagogy look like that teaches students to move between frictionless generation and irreversible commitment? I do not know yet. But I know both are needed, and I am beginning to see a glimpse of what that might look like. * * * The tools raise the floor for everyone. They do not raise the ceiling. The ceiling is still where it always was: what you actually have to say, and how much you are willing to risk in saying it precisely. The machine does not know what it made. You are responsible for knowing. That responsibility is not diminished by sharing the making. It is concentrated by it. --- # Intelligence in Deployment > Most AI systems that appear intelligent are not. They perform without closing the loop — and that distinction is the one that matters in the real world. - Author: Anieyrudh R (Singapore University of Technology and Design (SUTD)) - Published: 8 Apr 2026 - Type: Paper · Domain: AI Systems - Canonical page: https://bldg0.com/writing/intelligence-in-deployment A framework for judging real-world AI systems. Intelligence in deployment is task efficacy under real constraints, with uncertainty managed to a level sufficient for the task, sustained over time, with some degree of transfer across tasks or domains. ## Core Thesis AI should be judged by what it can reliably accomplish in deployment, not by whether it looks intelligent in isolation. A system that passes benchmarks, impresses in demos, and collapses under real constraints is not intelligent. It is performing. Working definition **Intelligence in deployment** = task efficacy under real constraints, with uncertainty managed to a level sufficient for the task, sustained over time, with some degree of transfer across tasks or domains. This framing is deliberately narrow. It excludes systems that reason well but cannot act, systems that act but do not learn, and systems whose competence does not survive contact with the messiness of the real world. What remains is a tight criterion: did the system get the job done, under the constraints that actually exist, despite the uncertainty that is always present, more than once, and can it do it somewhere else? ## The Loop Deployed intelligence is not a single act of reasoning. It is an ongoing cycle in which the system forms a workable picture of the world, anticipates consequences, commits to action, observes what actually happened, and becomes better calibrated for the next round. The cycle has five stages. Each exists because removing it breaks the loop — and the break is specific. Uncertainty is not a separate element that enters the loop at one point. It is present at every stage — in the same way that uncertainty is present in every physical measurement. Sensors are noisy, incomplete, delayed, and often only indirect proxies for what actually matters. Models of cause and effect are approximations. Decisions are commitments under imperfect knowledge. Feedback reveals discrepancies between prediction and reality. Intelligence lies not in eliminating uncertainty, but in managing it well enough to act effectively toward the objective. 1. **State inference.** The system turns raw, incomplete, noisy inputs into a usable internal picture of the world. It acts on its representation of reality, never on reality directly. This representation is always a best current estimate under incomplete access, with residual ambiguity — not the world as it is. The question is not whether uncertainty exists, but whether it is small enough for the task. _Kaelbling, Littman & Cassandra (1998)_ 2. **Prediction.** The system forms a cause-and-effect hypothesis: if the world is like this, and I do this, then this is what will likely follow. This is conditional reasoning under uncertain state and uncertain dynamics. _LeCun (2022)_ 3. **Decision.** The system commits. It selects a course of action and executes it in the world. This is where uncertainty becomes trade-off: do I know enough to act now? Is the cost of delay worse than the cost of acting imperfectly? 4. **Feedback.** The system encounters reality again. It observes the consequences of its action — whether its state estimate was off, whether its causal hypothesis was wrong. Feedback does not introduce uncertainty. It reveals and reshapes it. _Ashby (1956)_ 5. **Improvement.** The system uses accumulated feedback to become better over time — not only more accurate, but better at knowing what uncertainty matters, ignoring uncertainty that does not, and making better trade-offs under constraints. > Sensing is not yet intelligence. A system starts to become genuinely intelligent when it can move beyond raw input toward predicting consequences — under uncertainty — in a way that improves task completion. signal flowstage boundaryintelligence threshold The loop as a living system. Signals flow through five stages. The transition from sensing to prediction — marked as a phase change — is where intelligence begins. ## The Threshold Claim The strongest philosophical commitment of this framework is that intelligence does not begin at perception. A sensor-rich system is not automatically more intelligent. A system with many inputs may still perform poorly if it cannot compress, interpret, or use those inputs for action. Intelligence begins at prediction — specifically, at conditional causal reasoning of the form: _if the world is like this, and I do this, then this is what will likely follow._ This is not generic forecasting. It is not pattern matching. It is the formation of an interventionist hypothesis. This is where the framework diverges most clearly from the perception-heavy tradition in AI, which treats rich representation as a proxy for intelligence. It also diverges from the reward-is-enough position (Silver, Singh, Precup & Sutton, 2021), which treats reward maximisation as the single sufficient basis. The implication for engineering is direct: the real question is not _how much does the system sense?_ but _does it have enough task-relevant information to act effectively?_ Intelligence is partly a question of information sufficiency for control, not raw perceptual richness. ## The Diagnostic Structure The loop is not only descriptive. It is diagnostic. A system can fail at each stage, and the failures have distinct signatures. This is what makes the framework useful as an engineering tool rather than merely a philosophical taxonomy. A system that senses richly but infers poorly has an abstraction problem. A system that infers well but predicts weakly has a modelling problem. A system that predicts well but selects poor actions has a decision problem. A system that acts well but ignores feedback has a closure problem. A system that corrects locally but does not generalise has a transfer problem. These are not abstract categories. They correspond to different engineering interventions. The question shifts from _is this system intelligent?_ to _where in the loop is the bottleneck?_ bottleneck stagesignal attenuatedsignal flowing Five failure modes. Each bottleneck produces a distinct signature — and demands a different engineering intervention. ## The Evaluation Principle The diagnostic structure points toward a compact evaluation principle. Deployed intelligence can be assessed along four dimensions, and these dimensions are ordered. Design principle **Intelligence in deployment** = task efficacy × performance under real constraints × reliability over time × breadth of transfer. **Task efficacy** is the grounding term. Can the system get the job done? If the objective is not met, the rest is secondary. **Real constraints** is the deployment filter. Can it do so with the actual sensing, compute, time, and safety limits of the real world? **Reliability over time** is the consistency dimension. Can it do this repeatedly, not just once? Temporal consistency separates genuine capability from luck, overfitting, or narrow memorisation. **Breadth of transfer** is the generality dimension. Can the same underlying loop operate across different tasks, environments, or domains? ## Domain Application The loop stays structurally similar across domains. What changes is the meaning of state, action, feedback, and improvement. Domain State inference Prediction Decision Feedback Improvement Education Learner understanding Effect of intervention Hint, exercise, pacing Response, retention Refine tutoring Physical Object pose, force Effect of movement Action sequence Sensor measurement Refine control Software Status from logs Effect of config change Execute operation Errors, latency Refine behaviour Business KPIs, signals Effect of allocation Recommend, trigger Revenue, churn Update models Architecture Site, user, spatial Effect of arrangement Layout, material Simulation, use Refine design logic Economics Prices, sentiment Effect of policy Policy intervention Inflation, growth Update forecasts Structural invariance across six domains. The loop holds; the semantics change. ## Deep Walkthrough: Physical Robotics Physical robotics makes every stage of the loop concrete. Consider a manipulation system tasked with picking novel objects from an unstructured bin — a common warehouse application. **State inference** operates on depth-camera point clouds and RGB images. The system must segment individual objects from a cluttered pile, estimate their pose, and infer graspable surfaces — all from partial, occluded, noisy data. Bohg et al. (2017) demonstrated that perception and action are tightly coupled here: the system may need to nudge an object to reveal a graspable face. **Prediction** is where the system crosses the intelligence threshold. Given its inferred scene state and a candidate grasp pose, it must predict: will this grasp succeed? Will the object slip? This is conditional causal reasoning under physics. **Decision** involves trade-offs that cannot be reduced to maximising grasp probability. The system must balance grasp success likelihood against cycle time, collision risk, the probability of damaging the object or the gripper, and the value of an exploratory action versus a committed pick. **Feedback** is immediate and physical. The force-torque sensor registers whether the object was acquired. The vision system confirms whether the bin state changed as predicted. Each failure is diagnostic data for a specific stage of the loop. **Improvement** is what separates a deployable system from a demo. Over thousands of picks, the system should become better calibrated. A system that picks well once may be well-engineered or lucky. A system that picks better on its ten-thousandth cycle than its first is showing deployed intelligence. task efficacy (grounding)constraint tolerancereliability & transferdrag to rotate The evaluation space. Task efficacy is the grounding plane — without it, the other dimensions are irrelevant. Transfer is the vertical axis, the stretch toward generality. ## Positioning Against the Literature **Bounded rationality.** Russell and Subramanian (1995) established that agents should be judged by how well they achieve goals under computational limits. This framework shares that commitment entirely. Where it diverges is in making the internal loop itself explicit and diagnostic. **Partial observability.** The POMDP tradition (Kaelbling, Littman & Cassandra, 1998) formalised the separation between hidden world state, observation, inferred belief, and action under uncertainty. This framework absorbs that structure as the foundation of stage one. **Cybernetics.** Ashby (1956) placed feedback, adaptation, and environment-system coupling at the centre of intelligent regulation. This framework is a direct descendant. **World models.** LeCun (2022) argued that agents need internal models that predict consequences, support planning, and enable action. The overlap on the centrality of prediction is very strong. **Active inference.** Da Costa et al. (2020) and Parr, Pezzulo & Friston (2022) couple perception, prediction, and action through a generative model and uncertainty minimisation. The overlap on inference, prediction, action, and continual correction is substantial. * * * If this framework is accepted, it suggests a reorientation. The question worth asking is not whether a system is intelligent, but where in the loop it breaks — and whether the uncertainty at that stage is being managed well enough for the task at hand. The first question is a debate. The second is an engineering problem with a path toward resolution. Build systems that close the loop. That is where deployed intelligence lives. ## References - Ashby, W.R. (1956). _An Introduction to Cybernetics._ - Bohg, J. et al. (2017). Interactive Perception: Leveraging Action in Perception and Perception in Action. - Da Costa, L. et al. (2020). Active Inference and Reinforcement Learning. - Kaelbling, L.P., Littman, M.L. & Cassandra, A.R. (1998). Planning and Acting in Partially Observable Stochastic Domains. - Kroemer, O. et al. (2019). A Review of Robot Learning for Manipulation. - LeCun, Y. (2022). A Path Towards Autonomous Machine Intelligence. - Morris, M.R. et al. (2024). Levels of AGI: Operationalizing Progress on the Path to AGI. - Parr, T., Pezzulo, G. & Friston, K.J. (2022). _Active Inference._ - Russell, S.J. & Subramanian, D. (1995). Provably Bounded-Optimal Agents. - Silver, D., Singh, S., Precup, D. & Sutton, R.S. (2021). Reward Is Enough. --- # The New Nature of Work > Why AI Dissolves the Contract and What Takes Its Place - Author: Anieyrudh R (Singapore University of Technology and Design (SUTD)) - Published: 29 Mar 2026 - Type: Essay · Domain: Work × AI - Canonical page: https://bldg0.com/writing/the-new-nature-of-work AI has stripped away the scaffolding of busy-work that made task completion look like contribution. What remains is the question of what humans are actually for — and the answer is more interesting and more demanding than most organisations are prepared for. There is a poem by Jessie B. Rittenhouse that has always unsettled me: > _I bargained with Life for a penny,_ > _And Life would pay no more,_ > _However I begged at evening_ > _When I counted my scanty store._ > > _Life is a just employer,_ > _He gives you what you ask,_ > _But once you have set the wages,_ > _Why, you must bear the task._ > > _I worked for a menial's hire,_ > _Only to learn, dismayed,_ > _That any wage I had asked of Life,_ > _Life would have willingly paid._ Most people read this as a poem about ambition. I read it as a warning about assumptions — specifically, the assumptions we carry into work without ever examining them. The assumptions that quietly set our wages far below what life would have willingly paid. AI has just made those assumptions very expensive to hold. ## The Question Nobody Is Asking When people talk about AI and the future of work, the conversation almost always orbits the same anxious question: _which jobs will disappear?_ It is the wrong question. The more important question — and the one most teams are not asking — is this: _if a human's primary value is getting things done, why do you need the human at all?_ That question deserves to land. Sit with it. Because if the answer to "why is this person on the team" is some version of task completion — they write the code, they run the analysis, they produce the output — then you are already describing something AI can now do. And does better. Faster, cheaper, without ego, without sick days, without the friction of misaligned expectations. The tools exist. They are already in use. The question is no longer theoretical. This is not a threat to be managed. It is a clarification to be embraced. AI has stripped away the scaffolding of busy-work that made task completion look like contribution. What remains is the question of what humans are actually _for_ — and the answer, I want to argue, is more interesting and more demanding than most organisations are prepared for. ## Two Definitions of Work When people say _work_, they usually mean employment: a relationship where a person provides labour to another for payment, governed by a contract. This transactional model is where _work-life balance_ comes from. If a third of your waking life is spent living according to an arbitrary set of words you signed under financial pressure, liberation is not a want — it is a biological need. The contract is the cage, and the weekend is the parole. But there is a second definition, closer to the physics sense: any effort directed toward an intended outcome. No contract, no employer, no compulsion — just a person moving toward something they have chosen. By this definition, work resembles freedom more than its opposite. The reason most people default to task completion as their measure of contribution is not laziness. It is the logical response to being measured by a contract. Contracts can only recognise what is quantifiable. So people optimise for the quantifiable — the deliverable, the output, the ticket closed. The harder-to-measure things — curiosity, judgment, the quality of your thinking under pressure, your ability to make the people around you better — get quietly deprioritised, because they do not appear on any report. In large organisations, this was manageable. Enough complexity, enough bureaucratic mass, enough coordination overhead that the unmeasured things could hide inside the structure. Contribution was diffuse and hard to audit. The facade held. AI removes the facade. Smaller, more agile teams are becoming the unit of competitive advantage — not because smallness is fashionable, but because AI now handles the volume work that used to require headcount: the research, the drafts, the analysis, the coordination overhead. What remains in a team of five has to justify its presence far more acutely than it did in a department of forty. You cannot hide behind a job title or a standing meeting when the throughput is being handled by something that does not attend either. What you actually bring — your judgment, your relationships, your willingness to push back when the direction is wrong — becomes visible in a way it never was before. AI has not created this problem. It has made it impossible to ignore. ## What Cannot Be Automated Here is what AI cannot do: it cannot care whether the work is pointed at the right thing. AI is extraordinarily capable within a given value frame. It will optimise, generate, synthesise, and execute with a speed that makes human task completion look quaint. But it has no access to the question of whether the frame itself is correct. It cannot tell you when the team is moving fast in the wrong direction. It cannot feel the discomfort of a decision that is technically defensible but fundamentally wrong. It cannot push back on the founder who has fallen in love with their own idea. That is what humans are for. Not the output. The _judgment_. And judgment — real judgment, the kind that surfaces under pressure — does not emerge from individual intelligence. It emerges from friction. Specifically, from the friction of people reasoning together, disagreeing, misaligning, and then realigning around something truer than what any one of them held at the start. Most teams mistake communication for collaboration — keeping people informed, running alignment meetings, maintaining the appearance of shared direction. Real collaboration is something harder and rarer: people who care about different things being forced to reason together, to push back, to sit with disagreement until something truer emerges. It is where value alignment actually happens. It is where the chaos of real innovation is born. AI changes the stakes of this considerably. It compresses timelines, accelerates output, and expands the scope of what a small team can attempt — not just faster in the narrow sense, but broader and deeper than was previously possible. The same force that makes AI extraordinary also makes misdirection costly. A team moving in the wrong direction at human pace loses weeks. The same team, AI-amplified, can lose months in the same timeframe, having built further and faster down the wrong path. Which means the question of whether the work is pointed at the right thing becomes more critical, not less. And that question does not get answered in a document or a planning session. It gets answered through the continuous friction of people who are present, invested, and willing to be wrong in front of each other. Someone says a direction feels off. Someone else questions an assumption that everyone had quietly accepted. A third person connects it to something they noticed last week. These micro-corrections accumulate. The team's understanding of what matters — and why — gets stress-tested in real time, against reality, by people who care enough to say something. That is the feedback loop. Not the weekly standup. Not the project management tool. The constant, low-level friction of genuine collaboration. Remove it, and AI amplification accelerates in whatever direction was set at the last formal checkpoint. Speed, without that friction, becomes the enemy. ## Why This Requires Presence Remote work made sense in a particular era — one where human value was primarily measured by task completion, and tasks could be clearly scoped, handed over, and delivered asynchronously. In that model, presence was largely overhead. You could eliminate it without losing much that mattered. That era is over. And remote-first culture, which many organisations have mistaken for progressiveness, is quietly optimised for exactly the kind of work AI now does better. Decisions — real decisions, the kind that shape the direction of a team — do not happen in scheduled meetings. Cohen, March, and Olsen's garbage can theory describes organisations not as rational problem-solving machines but as loosely coupled streams of problems, solutions, people, and choice opportunities that collide, largely by proximity and chance. The implication is uncomfortable: the right problem finds the right solution not through formal process, but through the kind of unplanned collision that only physical presence makes possible. The hallway conversation that surfaces a concern nobody had thought to document. The shared lunch where two people realise they are solving the same problem from opposite ends. When an organisation is healthy, catching-up meetings are rare — not because people are well-informed, but because they are present enough that the collisions are happening continuously. Stephen Barley's research on how technology enters organisations makes this even more concrete. His finding was not that technology changes tasks — that much is obvious. It is that technology restructures the interactions between people: who depends on whom, who holds which knowledge, where authority actually sits. When CT scanners arrived in hospitals, radiologists did not simply get a better tool. Technicians suddenly held expertise that radiologists needed. Power shifted. The topology of the team reorganised. Remote-first culture, compounded by AI, performs the same restructuring — but toward shallowness. High-frequency, low-depth exchanges that simulate collaboration without generating it. The garbage can keeps moving, but the meaningful collisions stop happening. The friction of physical co-presence is not inefficiency to be engineered away. It is the medium through which teams develop the tacit, embodied understanding of each other that makes genuine collaboration possible. You cannot build the kind of relationship that spurs innovation over Telegram. You might sustain it in short bursts. You cannot generate it. ## The Redesign Required None of this is an argument against remote work as a tool. It is an argument against remote-first as a philosophy — against treating physical presence as a legacy default to be minimised rather than a generative condition to be designed around. The redesign that is actually required is more fundamental. Teams need to reconceive what they are for. The leader's role is no longer to assign tasks and monitor completion. It is to define the values, the constraints, and the direction — and then create the conditions under which people can push back, reason together, and arrive at something better. Hierarchy does not disappear, but hiding behind it does. A leader who uses their position to avoid productive friction is now a liability. They are removing the very mechanism that keeps the team oriented. The individual's role is no longer to execute. It is to bring something that cannot be automated: genuine engagement with whether the work matters, and the willingness to say so when it does not. That requires judgment — and judgment is not a given. It is earned slowly, through the experience of being wrong about things that mattered and having to course-correct with consequence. In an age where AI can absorb and generate knowledge at scale, the person who has actually lived through failure, sat with it, and emerged recalibrated is genuinely rare. That is the scarcity worth designing around. The conglomerates of tomorrow will not look like today's MNCs — sprawling organisations built to manage complexity through headcount. They will be symphonies of small teams: AI teams, human teams, and human-AI teams, each nimble and deeply concentrated in competence, coordinating across a shared value framework rather than a shared org chart. Agile teams that communicate closely and share knowledge fluidly across units will move faster and think better than any large, slow-moving hierarchy. The MNC as a vehicle for employing large numbers of people doing moderately differentiated work will decline. But this does not mean fewer jobs — it means different ones. If the unit of value is what you bring to others, more roles appear, not fewer. They will be more niche, more specific, less interchangeable. The demand for people who bring something genuine and irreplaceable will grow, not shrink. That requires believing in the human spirit — not as sentiment, but as a structural bet. That people, when the scaffolding of busywork is stripped away, will rise to the actual challenge of mattering. That the capacity for genuine contribution is there, latent, waiting for a context that demands it. But that belief is not given freely. It has to be earned — in the quality of your thinking, in the honesty of your pushback, in your willingness to develop yourself rather than wait to be developed. Every day. * * * The wages are different now. The contract model set them low, and most people bore the task without ever asking what else life might have paid. AI has dissolved the contract as the measure of contribution. What takes its place is harder to quantify and far more worth pursuing. Life will pay what you ask. The question is whether you know what to ask for. --- # Agentic Presence > Maintaining Intellectual Agency in AI‑Mediated Learning - Author: Anieyrudh R (Singapore University of Technology and Design (SUTD)) - Published: 22 Nov 2025 - Type: Paper · Domain: Education × AI - Canonical page: https://bldg0.com/writing/agentic-presence - PDF: https://bldg0.com/papers/agentic-presence.pdf Students increasingly report experiences like, 'I don't feel like I learned anything,' after using AI to write essays, debug code, or complete design work. This paper introduces agentic presence -- the learner's felt sense of intellectual authorship -- as a framework for understanding and addressing this disconnect in AI-mediated learning. ## Abstract Students increasingly report experiences like, "I don't feel like I learned anything," after using AI to write essays, debug code, or complete design work. Their outputs look sophisticated, yet their felt sense of learning is thin. This paper introduces _agentic presence_—the learner's felt sense of intellectual authorship, comprising intentionality, discernment, and ownership—as a framework for understanding and addressing this disconnect in AI-mediated learning. I propose a two-component evidentiary lens for learning: phenomenological indicators of felt agency combined with performance over time, especially in the quality of explanation and discussion. Both components are necessary; neither alone is sufficient. By distinguishing foundational skills that must transfer to the learner (Type 1) from capabilities that may remain legitimately distributed with AI tools (Type 2), and by situating these decisions within pedagogical intelligence, the framework offers a conceptual foundation for designing and studying learning environments where AI is pervasive but human judgment remains essential. The paper is deliberately theoretical: it articulates constructs, distinctions, and design implications intended as a bedrock for future empirical and design-based research, rather than specifying ready-to-deploy assessment instruments. ## 1\. Introduction A student uses AI to draft an essay, debug code, or design a solution. The output is sophisticated—perhaps exceeding what the student could independently produce. Yet afterward, the student reports an unsettling experience: "I don't feel like I learned anything." This disconnect between task completion and learning acquisition marks a fundamental disruption in educational practice. Traditional education rests on a tacit assumption: cognitive work—wrestling with problems, constructing arguments, debugging logic—builds intellectual capacity. Students write essays not because essays matter intrinsically, but because composing arguments develops argumentative thinking. They solve problem sets because working through mathematical reasoning builds mathematical capability. The cognitive struggle itself constitutes the learning. AI severs this connection. Students can now produce outputs without engaging the cognitive processes those outputs were designed to develop. This creates what I term the _capability asymmetry problem_: when learning with a system more capable than oneself, the learner risks defaulting to a follower role, delegating intellectual work rather than developing intellectual capacity. This challenge emerges against a broader transformation in work. Organizations are shifting from role‑based structures (fixed jobs with established procedures) to task‑focused environments (dynamic problems requiring adaptive problem‑solving). Success in such environments depends less on possessing specific knowledge and more on meta‑capabilities: Can you frame the right questions? Can you evaluate quality and judge risk? Can you orchestrate collaboration between humans, AI tools, and other contributors? Current educational responses to AI are inadequate. Banning AI denies students experience with tools that are reshaping professional practice. Embracing AI uncritically risks intellectual dependency—facility with AI's interface mistaken for genuine capability. Detection‑focused approaches misframe the issue as policing rather than pedagogy. In this paper, I develop a theoretical framework for maintaining intellectual agency in AI‑mediated learning, centered on _agentic presence_—the learner's felt sense of intellectual authorship. Agentic presence comprises three interrelated dimensions: - **Intentionality:** directing inquiry rather than passively receiving AI outputs; - **Discernment:** critically evaluating AI‑generated content rather than accepting it uncritically; - **Ownership:** accepting epistemic responsibility for outcomes despite AI's contribution. The framework serves two roles. Conceptually, it refines how agency can be understood and studied in AI‑rich learning environments. Pedagogically, it guides the design of tasks and environments that cultivate each dimension while recognizing that agentic presence must ultimately be paired with evidence of performance over time. This is a theoretical and conceptual paper. I do not attempt to solve questions of assessment feasibility or offer validated instruments. Instead, I introduce agentic presence as a foundational construct and outline distinctions and design principles that can guide future research and practice in AI‑mediated higher education. ## 2\. Why This Matters: From Role‑Based to Task‑Focused Work ### 2.1 The Changing Nature of Work Work is reorganizing around tasks, projects, and outcomes rather than fixed roles. A "job" increasingly represents a collection of tasks shared between humans, AI tools, and external contributors. Individuals and organizations collaborate through problems to be solved, not through predetermined procedures to be executed. This transformation changes what makes workers valuable. Success depends less on "what I know" in the narrow sense—codified knowledge held in memory—and more on "how I work with AI and others"—how effectively I use whatever knowledge I have to guide systems and collaborators toward useful outcomes. Differentiating capabilities increasingly include: - framing the right questions; - choosing and combining appropriate tools; - assessing, evaluating, and judging quality, risk, and impact. These meta-capabilities are fundamentally about thinking: adaptive problem-solving when stable procedures do not exist or cannot be fully specified in advance. ### 2.2 Why Develop "Thinkers"? The framework in this paper addresses developmental learning—education aimed at building transferable intellectual capacity rather than only immediate task performance. This distinction matters because AI's impact differs across educational goals. In functional or vocational learning, success measures are external: Does the bread taste better? Does the diagnosis help the patient? Here, AI can serve as a permanent tool enhancing performance without undermining learning's purpose. But in developmental or academic learning, the outputs themselves (essays, problem sets, projects) have no intrinsic value. Their sole purpose is developing student capability. An essay written by AI produces the artifact but fails educationally—like a treadmill that moves you forward without building cardiovascular fitness. Why privilege "thinking" as the educational goal in AI-rich environments? At least three reasons matter: 1. **Transfer and adaptability.** In task-focused environments where problems and technologies change rapidly, students need intellectual frameworks that apply across contexts. Thinking skills—problem framing, pattern recognition, explaining and justifying decisions—are more transferable than memorized procedures. 2. **Epistemic responsibility.** Democratic participation requires citizens who can evaluate claims, detect manipulation, and form reasoned judgments. In information environments where AI can produce persuasive but incorrect or manipulative content, critical evaluation cannot be safely delegated. 3. **Self-directed learning.** Task-focused work requires continuous adaptation. Workers must be self-directed learners capable of identifying what they need to understand, and of pursuing that. ### 2.3 The Measurement Problem and Bloom's Inversion Traditional education has relied on proxies for learning: grades, test scores, essay quality, project quality. These proxies worked tolerably well because producing outputs generally required developing the underlying capability. AI breaks this correlation through what I call _Bloom's taxonomy inversion_. Bloom's taxonomy orders cognitive skills from simple to complex: remember, understand, apply, analyze, evaluate, create. Students have traditionally ascended this hierarchy by mastering lower levels before higher ones. Current AI tools invert this sequence for many tasks. Large language models can create sophisticated outputs (essays, analyses, designs) without users necessarily able to analyze or evaluate those outputs effectively. Students can now create without understanding. This inversion exposes a deeper measurement problem. Educators cannot directly access student understanding; they infer it from behaviors and products. When AI can generate those products, traditional proxies become unreliable. A polished essay no longer guarantees that the writer can defend its claims; an elegant proof no longer guarantees that the student could reconstruct the reasoning. The framework I develop does not solve the entire measurement problem. It does, however, propose a two‑component evidentiary standard for learning in AI‑mediated environments: - **Phenomenological evidence:** Did the learner experience intellectual authorship—agentic presence—while working? - **Performance‑over‑time evidence:** Does capability persist and generalize across contexts, especially as expressed in the quality of explanations and discussions? Both components are necessary; neither alone is sufficient. A student may feel agentic but be confidently wrong. Another may perform well at a single task with heavy AI support while lacking durable capability. Learning, in this view, is signaled when agentic presence and performance trajectories align. ## 3\. Agentic Presence: Construct Definition ### 3.1 What Is Agentic Presence? Agentic presence is the learner's felt sense of intellectual authorship over their learning processes and outcomes. It is a phenomenological quality—how it feels to learn in a way that one experiences as "really thinking." When students report that they "didn't really learn" despite completing a task, they are reporting the absence of agentic presence. ### 3.2 Relation to Existing Constructs The construct I call agentic presence builds on, but is distinct from, several established ideas in the learning sciences: - **Epistemic agency.** In sociocultural and computer‑supported collaborative learning research, epistemic agency describes learners' capacity to take initiative in shaping the knowledge practices of a community (e.g., deciding what counts as a good explanation or what problems to pursue). It emerges from traditions influenced by Vygotsky and activity theory, emphasizing participation in shared practices rather than isolated cognition. - **Self‑regulated learning (SRL).** In educational psychology, SRL refers to learners' ability to plan, monitor, and regulate their cognition, motivation, and behavior toward goals (e.g., Zimmerman's work). It emphasizes strategies such as goal setting, self‑monitoring, and self‑evaluation, with roots in social‑cognitive theory and information processing. - **Metacognition.** Originating with Flavell, metacognition refers to knowledge about one's own cognition and the regulation of that cognition (e.g., knowing that one does not understand, and choosing strategies to repair understanding). - **Sense of agency.** In cognitive science and philosophy of mind, the sense of agency refers to the experience of controlling one's actions and their effects in the world—feeling like "I am the one who caused this." Agentic presence is not simply a new label for any of these. It is a phenomenological construct specific to AI‑mediated learning that: - focuses on the felt authorship of intellectual work, not just behavioral markers of regulation or participation; - decomposes that experience into intentionality, discernment, and ownership; and - is explicitly linked to a two‑component evidentiary standard for learning (phenomenology plus performance over time). In this sense, agentic presence can be seen as a bridge between traditions: it borrows the concern for participation and responsibility from epistemic agency, the attention to self‑direction from SRL, the introspective focus from metacognition, and the phenomenological grounding from sense of agency—while adding the specific challenge of maintaining authorship when working with a capable, non-responsible AI partner. ### 3.3 Phenomenological Reports as Reflective Practice Reports of agentic presence constitute a specific form of reflective practice. Dewey distinguished between mere activity and reflective experience—activity tied to conscious examination of meaning and consequence. In many educational settings, reflection takes the form of affective or general prompts: "How did this make you feel?" or "What did you think about this experience?" Such reflection can be valuable but often remains disconnected from the actual cognitive work of learning. Phenomenological reports of agentic presence are outcome-tied and action-anchored. Rather than asking only how students felt, they ask: - What questions did I formulate? - What decisions did I make in steering the AI? - Where did I accept or reject AI outputs, and why? - At which points did I feel responsible for the direction and quality of the work? This grounds reflection in specific intellectual acts while retaining the phenomenological dimension—the felt experience of being in control, exercising judgment, and owning outcomes. Because learning lacks direct sensory feedback accessible to others, the learner's phenomenological experience becomes a crucial source of evidence. At the same time, self-reports are fallible; students misjudge their own learning. Thus, in this framework agentic presence is a necessary but not sufficient indicator of learning. It must be paired with evidence that performance improves and generalizes over time, especially as expressed in the quality of the learner's explanations and contributions to discussion. ### 3.4 Three Interrelated Dimensions Agentic presence comprises three interrelated dimensions, each existing on a gradient rather than as a binary: - Intentionality - Discernment - Ownership Students may exhibit high intentionality but moderate discernment, or strong discernment but weak ownership. The dimensions are compensatory—strengths in one can partially scaffold weaknesses in others—but robust agentic presence requires all three to develop over time. Appendix A provides a matrix of combinations. **Intentionality** is the experience of directing inquiry rather than passively receiving information. Students with intentionality formulate genuine questions from curiosity or confusion, set epistemic goals, and maintain a coherent trajectory that connects current inquiry to prior understanding and future applications. Intentionality distinguishes inquiry (pursuing questions to develop understanding) from information-seeking (collecting answers to complete assignments). **Discernment** is the experience of evaluating AI outputs rather than accepting them uncritically. It is closely related to the idea of taste in intellectual contexts—a cultivated sense of what counts as a good explanation, a plausible argument, or an appropriate design. Students with discernment detect when AI contradicts itself or produces plausible but flawed reasoning, compare alternatives, and recognize AI's limitations in context. **Ownership** is the experience of epistemic responsibility for outcomes despite AI's contribution. It is relational: it concerns how students position themselves relative to AI and to their learning communities. Students with ownership acknowledge AI's role transparently, stand behind decisions made during AI interaction, accept consequences as their own, and experience outputs as reflecting their intellectual values and commitments. These dimensions blur in practice. When a student decides "what kind of answer would be useful," they are simultaneously directing inquiry (intentionality), exercising preliminary quality judgment (discernment), and taking responsibility for what counts as "useful" (ownership). The distinctions are analytic, not experiential; phenomenologically, they form facets of a single experience of intellectual authorship. ### 3.5 Agentic Presence, Domain Knowledge, and Feedback Loops Agentic presence is orthogonal to current domain knowledge. A novice can exhibit strong intentionality, emerging discernment, and genuine ownership while still lacking sophisticated conceptual understanding. In experiential and self-directed learning traditions, such agentic engagement is precisely how knowledge grows: learners inquire, evaluate, and take responsibility even in the presence of uncertainty. Early in learning, one should therefore expect configurations such as "high ownership, emerging discernment, low domain knowledge," where students are actively and responsibly engaged yet frequently wrong. In the terms of the matrix in Appendix A, these learners look like "earnest but uncritical" agents whose stance is desirable but whose evaluation skills are still developing. From this framework's perspective, such students are not failing; they are where a developmental process should begin. Conversely, students with substantial background knowledge can operate with low agentic presence if they delegate direction, evaluation, and responsibility to AI. The goal of education, then, is not simply to increase knowledge but to align agentic presence and knowledge growth over time. In practice, this alignment happens through creation-driven feedback loops. It is by doing something in the domain—writing an argument, implementing code, designing a solution—that learners generate the feedback they need to refine their understanding. Each cycle of "attempt → outcome → feedback → revision" updates the learner's internal models. Agentic presence describes the stance learners take within these cycles: whether they experience themselves as the ones posing questions, interpreting feedback, and deciding what to change. AI can supercharge these feedback loops in at least two ways. First, it increases the speed of iteration: where students might previously have produced a single draft or solution in a week, they can now generate and revise multiple variants in a single session. Second, it broadens exposure to variation: AI can quickly produce a wide range of possible solutions, structures, or styles. Under conditions of agentic presence—where learners direct prompts, critically evaluate outputs, and take responsibility for revisions—this accelerated, varied exposure can deepen both discernment and domain knowledge: students learn more quickly "what works and what doesn't," and come to appreciate the process that produces better outcomes. I return to this idea in the acceleration hypothesis (§8.5), where I connect these rapidly repeated creation–feedback cycles to Bloom's taxonomy inversion and argue that AI's ability to generate "create-level" artifacts at speed can, under agentic conditions, accelerate the development of judgment rather than bypass it. ## 4\. Intentionality: Directing Inquiry ### 4.1 Definition and Theoretical Grounding Intentionality captures the learner's sense of directedness: experiencing oneself as actively guiding inquiry rather than passively receiving AI's output. The term comes from phenomenology, especially Husserl's analysis of consciousness as always "consciousness of something"—directed toward an object rather than a neutral recording device. In learning, intentionality means experiencing oneself as directing cognitive activity toward understanding. This view aligns with constructivist traditions in education. Piaget conceptualized learning as active construction of mental structures; Dewey emphasized inquiry as a cycle of problem, hypothesis, and experimentation; Papert's constructionism foregrounded learners building artifacts as a way of thinking. Across these perspectives, genuine learning requires that the learner own the activity—hence the centrality of intentionality. Crucially, intentionality is not the same as prompting. Prompting is a mechanical act of typing; intentionality is an epistemic act of willing. A student can type "Write an essay on X" with zero intentionality, functioning only as a relay between assignment and AI. Intentionality exists only when the student has a specific epistemic goal and uses the AI to bridge the gap between current understanding and desired understanding. It is the difference between asking for an answer and pursuing an inquiry. ### 4.2 Addressing Capability Asymmetry Capability asymmetry—AI's superior performance on many tasks—threatens to position students as followers. In human–human collaboration, research on group work and expertise shows that less capable participants tend to defer to more capable ones, sometimes to the point of social loafing (reducing effort when a more capable partner carries the load). Intentionality counters this by establishing the learner as an epistemic agent. Even when AI generates sophisticated outputs, intentionality means the student: - formulates the question AI addresses; - determines what kind of answer would be useful; - decides whether the response actually addresses the question; - identifies what to pursue next based on gaps revealed. The student retains control over the direction of inquiry even while lacking control over the AI's capability to respond. A useful analogy is the relationship between a doctoral student and the research literature: the literature vastly exceeds the student's current capacity, yet the student can still decide which articles to read and what questions to ask of them. ### 4.3 Cultivating Intentionality Pedagogically Intentionality can be cultivated through design moves such as: - **Question‑first protocols:** Students must articulate questions and rationales before using AI: What do I need to understand? Why? What would constitute a satisfactory answer? - **Epistemic goal-setting:** For each AI interaction, students state their learning objectives (e.g., "I want to understand the mechanism," "I want counterarguments"). - **Trajectory documentation:** Students maintain learning logs linking current inquiry to prior understanding and anticipated future use. - **Socratic protocols:** Before consulting AI, students predict, hypothesize, or attempt answers themselves. - **Reflective pauses:** Time is deliberately set aside to ask: What am I trying to figure out? Am I making progress? What should I do next? These design moves also depend on environmental structures: if a course's pacing and grading regimes punish time spent on reflection as "overthinking," intentionality is structurally discouraged. ## 5\. Discernment: Critical Evaluation ### 5.1 Definition and Theoretical Grounding Discernment is the experience of critically evaluating AI outputs rather than accepting them uncritically. It involves developing taste—a refined sense of quality, coherence, and appropriateness. In Kant's aesthetics, taste represented cultivated judgment that balances general principles with sensitivity to particulars. Similarly, intellectual discernment develops when learners meet many instances, notice differences in quality, and abstract criteria for judgment. Schön's analysis of professional practice adds another layer. He distinguished reflection-in-action—thinking while doing—from reflection-on-action—retrospective analysis. A skilled designer evaluates a sketch as they draw it, guided by tacit criteria. In AI-mediated learning, discernment functions both retrospectively and in real time: students evaluate AI outputs as they arrive, and adjust prompts, accept or reject suggestions, and seek alternatives accordingly. ### 5.2 Addressing Authority Asymmetry AI's sophistication can generate authority asymmetry: students treat AI as an infallible expert. This resembles, but intensifies, novice–expert dynamics in human instruction. A novice programmer may defer to a senior engineer's code, assuming that "they must be right." Students can similarly defer to AI, particularly when they lack confidence in their own understanding. Discernment enables learners to treat AI as a resource rather than authority. A student may not be able to produce perfect code, but can still: - notice when AI's output contradicts earlier explanations; - identify logic that "does not quite follow"; - compare alternative outputs and articulate why one seems better suited to the task; - recognize when AI has ignored important contextual constraints. This trajectory parallels how novices develop taste in any domain. Wine tasters learn to identify flawed wine before they could ever produce excellent wine themselves. Readers learn to recognize weak arguments before they can craft strong ones. With AI, repeated exposure to varied outputs, including errors and contradictions, can accelerate discernment—if students are encouraged to compare and critique. ### 5.3 Cultivating Discernment Pedagogically Discernment can be deliberately developed through tasks such as: - **Comparative evaluation:** Present multiple AI outputs varying in quality; students rank and justify rankings against explicit criteria. - **Deliberate error analysis:** Provide AI‑generated outputs containing crafted flaws; students identify and classify errors. - **Criteria articulation:** Students collaboratively derive quality criteria by analyzing good and bad outputs. - **Multiple generation requests:** Students request several alternative responses from AI for the same prompt and compare them. - **Justification requirements:** Students must explain why they accepted, modified, or rejected AI outputs. - **Progressive refinement documentation:** Students record how they iteratively improve AI outputs and why. Discernment is particularly sensitive to environmental and cultural norms. If a course rewards speed and correctness without valuing discussion of why an answer is good or bad, discernment has little room to grow. ## 6\. Ownership: Epistemic Responsibility ### 6.1 Definition and Theoretical Grounding Ownership is the experience of epistemic responsibility for outcomes produced with AI. It is not about provenance (who wrote the words) but about answerability (who stands behind the meanings). The notion of answerability is central in Bakhtin's dialogic theory: speakers are responsible for their utterances, not because they authored every word in isolation, but because they endorse what their utterances do in the world. In AI‑mediated work, students are answerable for the claims they submit, the decisions they make, and the designs they deploy, even when AI generated substantial portions of the work. Ownership means: - acknowledging AI's role transparently; - taking responsibility for choosing to trust or modify AI outputs; - accepting consequences when outputs are wrong or harmful; - connecting the work to one's own developing identity as a thinker or professional. Students can have strong ownership even when AI contributes heavily to the text or code, if they have genuinely evaluated and chosen to endorse the output. Conversely, they can lack ownership even over work they typed themselves, if they do not feel responsible for its content. ### 6.2 Addressing the Boundary Problem Traditional assessment regimes assume clear authorship boundaries: either the student wrote it or someone else did. AI complicates this picture. Students may legitimately collaborate with AI in planning, drafting, revising, and polishing. The question shifts from "Is this yours?" to "How did you think through this?" or "Where, in this process, did you exercise judgment?" In this context, decision points become more important than provenance. A student who prompts AI to generate an essay but then critically evaluates coherence, identifies gaps, requests revisions, and integrates insights has engaged meaningfully, even if many sentences remain AI‑authored. Another student might minimally tweak AI output and submit it without thought. Ownership differs dramatically between these cases, despite similar surface artifacts. ### 6.3 Cultivating Ownership Pedagogically Ownership can be supported by emphasizing transparency and responsibility: - **Process documentation:** Students document AI use: prompts, outputs, decisions ("I accepted this because… I rejected this because…"). - **Justification portfolios:** Students defend work in writing or orally, explaining key decisions and how they relate to learning goals. - **Revision analysis:** Students annotate how they modified AI outputs and why. - **Public articulation:** Students present work and answer questions without AI assistance. ## 7\. Integration and Collaborative Epistemology ### 7.1 Gradients, Compensation, and Development The three dimensions of agentic presence exist on gradients and can compensate for each other to a degree. For example: - A student may have high ownership and intentionality but low discernment: they drive inquiry and feel responsible, but are easily misled by plausible AI outputs ("earnest but uncritical"). - Another may have strong discernment and ownership but low intentionality ("reactive expert"): they can evaluate well and stand behind their choices, but tend to respond to whatever AI presents rather than driving inquiry. - A third may display curiosity (intentionality) but low discernment and ownership ("curious delegator"): they ask interesting questions but outsource evaluation and responsibility to the AI. These configurations are developmentally meaningful. Early in learning, "earnest but uncritical" is not a failure; it is a starting point. The task of pedagogy is to use strengths in one dimension to scaffold growth in others—for instance, using high ownership to motivate students to develop better discernment ("if I am responsible for this, I should care whether it's actually correct"). Critically, agentic presence is not identical to expertise. A novice can be highly agentic and yet often wrong; an expert can be correct while operating in a low‑agency mode with respect to AI. The educational target is not simply high correctness, but a trajectory in which agentic presence and capability grow together over time. ### 7.2 Human–Human Analogues: Group Work as Pre‑AI Precedent The dynamics described so far are not unique to AI. In traditional group work, students with lower agency often "ride along" with more agentic peers. A highly capable, high‑agency student frames the problem, decides on an approach, evaluates solutions, and writes up results; others primarily execute tasks or simply attach their names to the final product. This pattern is well‑documented in research on group work and social loafing: when individual contribution is hard to distinguish, some members exert less effort, relying on others to carry cognitive and practical load. AI makes a similar dynamic available in individual work: a student can now "ride along" with an indefinitely capable partner. Agentic presence extends naturally to this context. A group exhibits high agentic presence when its members collectively direct inquiry, evaluate outputs, and share responsibility for decisions. A group where one member sets all goals, makes all judgments, and carries ownership while others simply comply has low distributed agentic presence—even if the final artifact is excellent. AI does not create the problem of asymmetric agency; it scales and automates it. What was once an occasional group‑work issue now becomes a persistent possibility any time students can offload intellectual work to AI. This continuity suggests that insights from research on group work and epistemic agency can inform how agentic presence is cultivated in AI‑mediated tasks. ### 7.3 Collaborative Epistemology, Dialogism, and the Sounding‑Board Metaphor Taken together, intentionality, discernment, and ownership support a form of collaborative epistemology—the practice of generating knowledge through human–AI interaction while maintaining human epistemic responsibility. This view builds on work in distributed cognition, which shows that cognitive processes are often spread across people, tools, and representations rather than residing only "in the head." However, collaborative epistemology adds a crucial normative condition: although cognition may be distributed across human and non‑human elements, epistemic responsibility is not. AI systems contribute to the process by generating candidate meanings, but they cannot be bearers of epistemic responsibility because they have no commitments, no stake in the outcome, and no capacity for accountability. This position distinguishes human–AI collaboration both from human–human collaboration and from the use of "dumb" tools. In human–human collaboration, partners can share understanding and responsibility; both parties can, in principle, be addressed as knowers who can give and receive reasons. In contrast, calculators and similar tools perform well‑specified operations but do not generate new candidate interpretations or arguments; they compute but do not propose meanings. Current large language models sit between these cases. Like human collaborators, they can produce fluent explanations, arguments, and designs, and can sustain extended, context-sensitive dialogue. Unlike human collaborators, they lack understanding and cannot be answerable for what they generate. Unlike calculators, they do far more than implement fixed functions: they propose interpretations that invite uptake. Collaborative epistemology names this hybrid situation: humans and AI jointly produce candidate knowledge claims, but only humans can own and stand behind those claims. Bakhtin's notion of dialogism helps explain why this matters. For Bakhtin, meaning arises in dialogue among voices: each utterance is addressed to someone, anticipates a response, and carries with it a form of answerability. Speakers are responsible for what their utterances do in the world. AI outputs resemble utterances in their linguistic form, but they lack addressivity and answerability in the Bakhtinian sense. They are not "someone's" voice in the way a human utterance is; they do not anticipate a response, nor can they be held to account. In collaborative epistemology, the student is the one who re-voices AI-generated text as their own utterance in a human dialogue—toward a teacher, a peer, a client, or their future self. Meaning is not located in the AI's raw text but in how the human takes it up, modifies it, and addresses it to others. The dialogic relation that matters is ultimately between humans, even when mediated by AI. Within this framing, the sounding-board metaphor becomes more precise. A sounding board responds but does not initiate; it amplifies and colours sound without having its own musical intentions. Similarly, AI responds to prompts without setting epistemic agendas of its own. Protecting intentionality means preserving this asymmetry: the human voice must initiate and steer the dialogue. AI amplifies and colours ideas, but discernment is needed to distinguish signal from noise, recognizing that AI's contributions reflect training data and model architecture rather than understanding. Ownership, in Bakhtin's sense of answerability, requires that learners treat AI outputs as material to be appropriated and readdressed—not as final utterances that relieve them of responsibility. Pedagogically, this form of collaborative epistemology looks like students conducting a dialogue with AI whose endpoints are always human. A student working on an engineering concept, for instance, might (1) pose an initial question to clarify a mechanism, (2) ask AI to generate multiple explanations or design options, (3) critique those options, requesting revisions and counterexamples, and (4) compose a final explanation or design rationale addressed to a human audience. Throughout, AI supplies candidate meanings, but the student frames the questions, judges plausibility, and decides what to stand behind. ### 7.4 Dialogue Over Delegation: Active Curation Collaboration with AI can take the form of dialogue or delegation. In delegation, a student issues a single, broad prompt ("Write an essay on X") and accepts the first answer. Intentionality, discernment, and ownership are minimal. In dialogue, the student engages in sustained exchange, where each prompt is shaped by evaluation of prior outputs. For example: - **Elaboration sequences:** "Explain concept X." → "What is confusing is Y; clarify." → "Give an example." → "How does this connect to Z?" - **Challenge‑response sequences:** "Present the case for A." → "What is the strongest counterargument?" → "How would proponents respond?" → "Which side is more convincing and why?" - **Comparative evaluation sequences:** "Give three approaches." → "What are the trade‑offs between them?" → "Given my constraints, which is most appropriate and why?" In such dialogue, students must continually exercise intentionality (deciding what to ask next), discernment (judging responses), and ownership (choosing and justifying paths forward). Learning design that expects and rewards dialogue, rather than mere delegation, is more likely to foster agentic presence. ## 8\. Capability Types and Pedagogical Intelligence ### 8.1 The Scaffolding Paradox Traditional scaffolding theory (Wood, Bruner, & Ross) assumes that supports are gradually removed as learners internalize capabilities. However, AI tools—especially large language models—are not designed to fade. They will likely remain as a permanent layer of cognitive infrastructure in many domains. If education insists that all capabilities must be performed independently of AI to count as "learned," it risks preparing students for a world that no longer exists. Conversely, if education treats any capability that AI can perform as unnecessary to develop, it risks hollowing out the very foundations that make judgment possible. Resolving this paradox requires a principled distinction between different types of capabilities. ### 8.2 Type 1 Capabilities: Must Transfer Type 1 capabilities are foundational cognitive skills that must be internalized by the learner to enable effective judgment and agentic presence. They are prerequisites for intentionality, discernment, and ownership. The logic is simple: one cannot evaluate what one does not fundamentally understand. To judge whether an AI's argument is sound, the learner must possess at least rudimentary logical competence. To direct inquiry, the learner must have a mental model of the domain that makes certain questions meaningful and others irrelevant. Examples of Type 1 capabilities include: - basic logical reasoning and argument evaluation; - foundational domain concepts (e.g., conservation laws in engineering, core mechanisms in biology); - the ability to follow and reconstruct an explanation at an appropriate level; - framing meaningful questions and recognizing when a question has been answered. For Type 1 capabilities, AI can serve as a temporary scaffold but must eventually be faded in assessment. A minimal test is: if the AI disappears, can the learner still function at a basic level? For Type 1 capabilities, the answer must be yes. Crucially, any capability that underwrites agentic presence itself must be Type 1. If students lack basic logical tools, discernment collapses; if they lack any internal sense of epistemic responsibility, ownership collapses. The underlying skills that make intentionality, discernment, and ownership possible cannot be safely outsourced. ### 8.3 Type 2 Capabilities: Legitimate Distribution Type 2 capabilities are capabilities where cognition can remain legitimately and permanently distributed between human and machine. Here the goal is not independent performance, but orchestration expertise: the ability to use tools to achieve results that neither the human nor the AI could produce alone. Examples include: - large‑scale data synthesis and summarization; - generating boilerplate code, documentation, or routine text; - translating between technical and non‑technical registers; - stylistic polishing of prose. For Type 2 capabilities, AI is not a temporary scaffold but a permanent tool. Assessment should focus on how well the learner orchestrates the tool: Can they direct the AI to a better result than a novice user could? Can they adapt its output to context? Can they detect when the tool's limitations are being exceeded? These distinctions are not static. What counts as Type 1 versus Type 2 can shift over time and across disciplines. Here pedagogical intelligence—educators' judgment about which difficulties are still productive for learning—plays a central role. ### 8.4 Pedagogical Intelligence and the Type 1/Type 2 Boundary The term _pedagogical intelligence_ here extends Shulman's notion of pedagogical content knowledge: it is the practical wisdom educators use to decide what to teach, how, and to what end in a changing technological environment. Pedagogical intelligence includes the ongoing task of deciding which capabilities remain Type 1 and which can be treated as Type 2. These decisions are partly disciplinary (what does it mean to think like an engineer or designer?), partly ethical (what responsibilities should never be delegated?), and partly pragmatic (what is realistically needed for future work). Within this framework, one claim is non‑negotiable: > Epistemic responsibility—ownership—is always a Type 1 capability. No matter how sophisticated AI becomes, one cannot outsource responsibility for the consequences of decisions. Similarly, the basic logical and conceptual tools needed to exercise discernment and intentionality must remain Type 1. Other capabilities may shift categories as AI improves; these cannot. ### 8.5 The Acceleration Hypothesis: From Drudgery to Pattern Recognition A common objection to emphasizing discernment and orchestration is the bootstrapping problem: How can novices develop good judgment about AI outputs if they have not first struggled through execution themselves? Traditional models imply a linear trajectory: first master execution (e.g., writing code or proofs by hand), then develop meta-level judgment. Agentic presence suggests a parallel trajectory. The same Bloom's taxonomy inversion described in §2.3—AI's ability to generate "create-level" outputs for learners who may still be at "understand" or "apply"—is usually treated as a threat to assessment validity. But it also creates an opportunity: because AI can produce sophisticated artifacts so quickly, learners can move through full cycles of creation → evaluation → revision far more frequently than before. In a traditional course, a student might hand-craft one essay or one code solution per week. Each cycle of creation and feedback is slow. With AI, a student who maintains intentionality, discernment, and ownership can, in a single session: - generate multiple complete drafts or solutions; - compare them, noticing differences in structure, clarity, or robustness; - ask AI to adjust or recombine them in specific ways; and - reflect on how these changes affect quality. In other words, the frequency of full iteration loops increases dramatically. Instead of a single, slow ascent up Bloom's hierarchy to one instance of "create," students repeatedly encounter, critique, and reshape create-level artifacts. Under conditions of agentic presence, this repetition accelerates pattern recognition and the development of taste: students begin to see recurring features of strong and weak solutions across many examples—common failure modes in code, typical weaknesses in arguments, or hallmarks of clear explanation. This does not mean that execution no longer matters. Rather, it suggests a division of labor across time: - Type 1 capabilities at the execution level (basic logic, core concepts, minimal unaided performance) develop through a mix of traditional practice and carefully constrained use of AI as a scaffold; - judgment-oriented Type 1 capabilities (discernment in particular) develop through high-frequency cycles of evaluating and refining AI-generated creations. AI thus becomes a machine for accelerating intuition and taste if students remain authors of the process. Without intentionality, discernment, and ownership, Bloom's inversion simply produces more polished artifacts with no corresponding growth. With agentic presence, the inversion becomes a resource: rapid creation enables equally rapid cycles of evaluation and reflection, allowing Type 1 capabilities—especially discernment—to develop more quickly and broadly than they would through slow, solitary production alone. ## 9\. Designing Environments for Agentic Presence ### 9.1 Cognitive Load Redistribution Cognitive load theory distinguishes among intrinsic, extraneous, and germane load. AI radically redistributes these loads. Many tasks that once carried intrinsic and germane load for learners (e.g., low-level coding, mechanical derivations) can now be offloaded. At the same time, new germane loads emerge: - evaluating output quality; - identifying AI's limitations and failure modes; - maintaining coherent direction across extended interactions; - synthesizing across multiple AI outputs and other sources; - monitoring one's own understanding in the midst of AI assistance. Pedagogical intelligence requires continually asking: Which loads remain germane in an AI-rich environment? Some historically germane tasks may now be extraneous (e.g., laborious hand-calculation when the concept, not the arithmetic, is the learning target). Others remain essential as Type 1 foundations. New germane loads appear around coordination and judgment. ### 9.2 Process‑Focused Evidence and Performance Over Time Because AI undermines traditional artifact‑based proxies, the framework emphasizes process‑focused evidence and performance over time as conceptual indicators of learning. The two‑component evidentiary standard can be summarized as: - **Phenomenological evidence:** Does the learner report having directed inquiry, evaluated outputs, and owned decisions—agentic presence? - **Performance‑over‑time evidence:** Does the learner's performance improve and generalize across tasks, especially as reflected in the quality of explanation and discussion? This paper does not prescribe specific instruments; instead, it sketches the kinds of evidence that would align with the framework. For example: - **Process portfolios.** Students assemble traces of their interaction with AI (prompts, outputs, annotations, reflections) to make their thinking visible. - **Quality of discussion.** Over time, instructors can attend to changes in how students talk about problems: Do their questions become more precise? Do their explanations show deeper structure? Are they able to critique AI outputs in increasingly nuanced ways? - **Transfer tasks.** After AI‑supported work in one context, students tackle related but novel tasks where AI support is limited or structured differently. The key point is conceptual: evidence of learning in AI‑mediated environments must be longitudinal and process‑sensitive. Agentic presence should correlate with richer trajectories of contribution and explanation, not merely with isolated correct answers. ### 9.3 Environmental Affordances and Space for Reflection Agentic presence is not only an individual capacity; it is also shaped by environmental affordances—what the learning environment makes easy or hard to do. The concept of affordances originates in ecological psychology (Gibson), where it describes the possibilities for action that an environment offers. In educational settings, structures, schedules, and norms afford or constrain reflection, questioning, and experimentation. Some environments systematically make agentic presence difficult: courses packed with tasks and deadlines that leave no time for genuine reflection; assessment regimes that reward speed and minimal compliance; norms that label questioning AI outputs or revisiting problems as "overthinking." In such environments, students receive a clear message: do not slow down; do not question too much; just get to the answer. Intentionality, discernment, and ownership become costly; effortless delegation becomes rational. Designing for agentic presence therefore requires creating temporal and normative space for reflection and critique. This includes: - building scheduled pauses into activities to interrogate AI outputs; - explicitly valuing time spent comparing, critiquing, and re‑framing; - modeling and legitimizing skeptical engagement with AI as a professional norm rather than resistance. Pedagogical intelligence, in this expanded sense, involves not only deciding what remains Type 1, but also shaping environments so that exercising agentic presence is both possible and expected. ### 9.4 Competency‑Based Evaluation and Preparation for Future Work Finally, the framework supports a shift from task completion to competency demonstration. Instead of asking primarily whether students have produced correct outputs, educators can ask: - Can this student frame meaningful, contextually relevant questions with and without AI? - Can this student evaluate and adapt AI outputs in light of domain constraints? - Can this student orchestrate human and machine contributions while retaining ownership? These competencies—intentionality, discernment, and ownership in practice—align closely with what task‑focused, AI‑mediated work demands. Students who develop agentic presence become self‑directed learners capable of navigating unstable environments, rather than merely proficient operators of current tools. ## 10\. Open Questions and Limitations This paper makes strong conceptual claims but deliberately remains agnostic about several important issues. **The cognitive cost of agency.** Maintaining agentic presence is cognitively expensive. Delegating to AI is efficient; actively directing, evaluating, and owning is effortful. For students—and institutions—accustomed to valuing efficiency, the friction of agency can feel like wasted effort. A major challenge for future work is to understand how to justify and sustain this cognitive investment, both to learners and to systems that reward throughput. **Empirical validation and measurement.** I have proposed agentic presence as a construct and outlined a two‑component evidentiary standard, but I have not specified instruments or addressed issues of construct validity or reliability. Future research will need to develop methods—qualitative, quantitative, and mixed—that can capture phenomenological reports, process traces, and performance trajectories in ways that are both rigorous and feasible. **Affective dimensions (bracketed here).** Agentic presence, as defined here, focuses on the phenomenological experience of authorship and the cognitive capacities that support it. I have deliberately bracketed students' emotional states—anxiety, boredom, motivation—even though these undoubtedly interact with agentic presence in practice. For example, fear may push students toward delegation, while overconfidence may fuel uncritical ownership. Examining these interactions is an important direction for future work but lies beyond the scope of this initial theoretical formulation. **Institutional and equity constraints.** Process‑sensitive and longitudinal approaches to evidence demand time and expertise from instructors and institutions. They may be difficult to scale in large, resource‑constrained systems. Moreover, access to AI tools, prior experiences with schooling, and cultural norms of deference may shape who develops agentic presence and who does not. The framework does not yet address these equity issues; it provides, at best, a language for asking sharper questions about them. Despite these limitations, I argue that articulating agentic presence and the associated distinctions is a useful starting point. Without such constructs, efforts to integrate AI into education risk defaulting either to nostalgia (re‑creating pre‑AI conditions) or to naïve technophilia (equating tool use with capability). ## 11\. Conclusion This paper addresses a fundamental challenge of AI‑mediated learning: AI systems can complete traditional learning tasks without students developing the underlying capabilities those tasks were designed to cultivate. Rather than framing this challenge as a matter of detection or prohibition, I have proposed a pedagogical and conceptual response. The central construct, agentic presence, names the learner's felt sense of intellectual authorship and decomposes it into intentionality, discernment, and ownership. These dimensions articulate what it means to "really think" with AI, rather than merely operating it. They provide both a diagnostic lens—helping educators see where learning may or may not be happening—and a design target—guiding the construction of tasks and environments that cultivate agency. The framework further distinguishes between Type 1 capabilities that must transfer to learners (including the foundations that underwrite agentic presence itself) and Type 2 capabilities that may remain legitimately distributed with AI tools. It situates these distinctions within pedagogical intelligence—the evolving practical wisdom of educators about which difficulties remain productive in an AI‑rich world. Finally, the paper argues for a two‑component evidentiary standard for learning: phenomenological reports of agentic presence paired with performance over time, especially as expressed in the quality of explanations and discussions. Learning, in this view, is not fully captured by static artifacts; it is seen in trajectories of engagement and responsibility. AI does not force education to choose between technological relevance and intellectual development. The challenge is to cultivate agency through AI collaboration rather than despite it—to design learning experiences in which students come to experience themselves as the authors of their own thinking, even when thinking with powerful machines. ## References - Bakhtin, M. M. (1993). _Toward a philosophy of the act_ (V. Liapunov, Trans.). University of Texas Press. (Original work written 1919–1921) - Bloom, B. S. (Ed.). (1956). _Taxonomy of educational objectives: The classification of educational goals. Handbook I: Cognitive domain_. Longmans, Green. - Dewey, J. (1938). _Experience and education_. Macmillan. - Flavell, J. H. (1979). Metacognition and cognitive monitoring: A new area of cognitive–developmental inquiry. _American Psychologist, 34_(10), 906–911. - Gibson, J. J. (1979). _The ecological approach to visual perception_. Houghton Mifflin. - Husserl, E. (1970). _Logical investigations_ (J. N. Findlay, Trans.). Routledge & Kegan Paul. (Original work published 1900–1901) - Kant, I. (2000). _Critique of the power of judgment_ (P. Guyer, Ed., P. Guyer & E. Matthews, Trans.). Cambridge University Press. (Original work published 1790) - Papert, S. (1980). _Mindstorms: Children, computers, and powerful ideas_. Basic Books. - Piaget, J. (1952). _The origins of intelligence in children_ (M. Cook, Trans.). International Universities Press. - Pollitt, A. (2012). The method of comparative judgement. _Assessment in Education: Principles, Policy & Practice, 19_(3), 281–300. - Salomon, G., & Perkins, D. N. (1989). Rocky roads to transfer: Rethinking mechanisms of a neglected phenomenon. _Educational Psychologist, 24_(2), 113–142. - Schön, D. A. (1983). _The reflective practitioner: How professionals think in action_. Basic Books. - Shulman, L. S. (1986). Those who understand: Knowledge growth in teaching. _Educational Researcher, 15_(2), 4–14. - Sweller, J. (1988). Cognitive load during problem solving: Effects on learning. _Cognitive Science, 12_(2), 257–285. - Vygotsky, L. S. (1978). _Mind in society: The development of higher psychological processes_. Harvard University Press. - Wood, D., Bruner, J. S., & Ross, G. (1976). The role of tutoring in problem solving. _Journal of Child Psychology and Psychiatry, 17_(2), 89–100. - Zimmerman, B. J. (2002). Becoming a self‑regulated learner: An overview. _Theory Into Practice, 41_(2), 64–70. ## Appendix A. Agentic Presence Component Matrix Intentionality Discernment Ownership Result Example Manifestation High High High Full agentic presence Student pursues their own questions, critically evaluates AI responses, and stands behind the integrated work as reflecting their understanding. High High Low Brilliant but uncommitted Student asks sophisticated questions and evaluates well, but attributes success/failure to AI rather than owning decisions. High Low High Earnest but uncritical Student drives inquiry and owns work but accepts AI outputs without sufficient evaluation, confidently owning flawed understanding. Low High High Reactive expert Student evaluates AI outputs well and owns decisions, but rarely initiates new lines of inquiry—acts as a sophisticated consumer rather than an independent thinker. High Low Low Curious delegator Student asks interesting questions but does not evaluate or take responsibility for AI outputs—hands off thinking at the earliest opportunity. Low High Low Critical observer Student can critique AI quality but does not direct inquiry or own outcomes—stands at the edge of the work, commenting rather than participating. Low Low High Misplaced ownership Student claims responsibility for work they neither directed nor evaluated—performs ownership without substance. Low Low Low Complete delegation Student remains entirely passive—AI does the work; the student submits it without intellectual engagement. ---