// SLIDE 01 — HOOK

CHAPTER 12 — CAPSTONE: THE SUPERVISED AGENTIC PROJECT

GenAI assessment research has identified a consistent failure mode: evaluating agentic or AI-assisted work by the qualit

Chapter 12 — Capstone: The Supervised Ag — Microsoft Research's study on generative AI and critical thinking (CHI 2025) documents the mechanism: self-reported redu

NARRATION

GenAI assessment research has identified a consistent failure mode: evaluating agentic or AI-assisted work by the quality of the output alone (arXiv, 2025, "Navigating the New Landscape"). A polished deliverable can hide unverified claims that happen to sound plausible, data inputs that were out of scope or outdated, reasoning steps the agent compressed or skipped, actions the agent took that no human approved, and omissions the agent made silently.

// SLIDE 02 — THE TEN ARTIFACTS

THE TEN ARTIFACTS.

Project brief.→Data boundary.→Action surface map.→Agent plan.→Approval gates.

NARRATION

A supervised agentic project produces ten artifacts. Together they answer: what was the task, what was the agent allowed to do, what happened, and how do I know the output can be trusted? Project brief. What is the goal? What is the deliverable? Who is the audience? What would success look like, and what would failure look like? A brief that cannot answer these questions is not ready to delegate to an agent.

// SLIDE 03 — A WORKED EXAMPLE: RESEARCH BRI

A WORKED EXAMPLE: RESEARCH BRIEF

Most likelythe agent synthesizes claims not directly supported by the specific papers.

Secondthe brief omits one paper's central disagreement with the others.

Thirdcitations are formatted plausibly but do not match actual page locations in the originals.

Fourththe agent offers "relevant context" from its training data rather than sticking to the three sources.

NARRATION

The project: summarize three recent peer-reviewed papers on AI governance published in the last two years, comparing their recommendations for organizational oversight. Project brief. Goal: a three-page comparative brief. Deliverable: a structured document with source citations. Audience: a team making policy decisions. Success: accurate, cited, clearly comparative. Failure: uncited claims, outdated papers, fabricated details, missing differences between the papers' positions.

// SLIDE 04 — TRACKS

TRACKS

Chat-only.→Cowork workflow.→Claude Code workflow.→MCP capability review.

NARRATION

The capstone structure works across different access levels. The ten artifacts are the same regardless of track. The tools change; the supervision does not. Chat-only. Use Claude AI to work through the capstone structure in conversation. Produce all ten artifacts as text documents. This track practices the thinking without requiring Cowork or Code access.

// SLIDE 05 — THE CLOSING ARGUMENT

THE CLOSING ARGUMENT.

You started this book with the observation that a polished artifact announces nothing about whether the work behind it w

The Closing Argument — The book's argument has been consistent: agentic AI is delegated action under constraints, and the capable human is the

NARRATION

You started this book with the observation that a polished artifact announces nothing about whether the work behind it was sound. An agent can produce fluent text, passing code, formatted tables, and plausible-sounding citations. None of that is evidence that the output is correct, appropriately scoped, or worth acting on.

// SLIDE 06 — THESIS

THE CORE CLAIM.

You started this book with the observation that a polished artifact announces nothing about whether the work behind it w

The book's argument has been consistent: agentic AI is delegated action under constraints, and the capable human is the

NARRATION

// SLIDE 07 — CLOSE

ASK THE QUESTION. APPLY THE FRAMEWORK.

THE TEN ARTIFACTS//A WORKED EXAMPLE: RESEARC//TRACKS

Claude Agentic AI · Ch.12 · Chapter 12 — Capstone: The Supervised Agentic Project

NARRATION

That is the framework. Claude Agentic AI, chapter 12: Chapter 12 , Capstone: The Supervised Agentic Project. The patterns are now in place. Apply them.