Guide · First edition, October 2026
Contents
Full chapter text is on the web. This list mirrors the Guide. Prefer/guide/ for reading with previous and next links.
- —Foreword: why I ran the experimentsFindings in brief, and how findings differ from recommendations.
- 01Facts, checks and upkeepTerms, the evidence in outline, and the worked example the rest of the book follows.
- 02Briefs, plans and specificationsDecide what the change is, how you will know it is done, and what must not change, before an agent writes code.
- 03The right facts, not every factThe main evidence chapter: generated context on a readable repository, stale dependency knowledge, and facts across a repository boundary.
- 04One verifiable change at a timeHand an agent one change you can verify at a time; the quality hangs on step size and the verify step.
- 05Proof, not promisesExecutable checks and an independent judgement of the work, not only the agent’s account of it.
- 06The harness and the friction loopAgents trip on the same few environment problems; put fixes in the harness around them, not only in prompts.
- 07Parallel agentsDeliberately short: none of my experiments measured parallel working as such, so this is judgement and other people’s published practice.
- 08Review, merging and riskReview by the consequence of being wrong: reversibility, blast radius and verification strength.
- 09Tools and honest measurementPrefer boring tools with clear errors, and measure set-up changes with repeated runs rather than single anecdotes.
- 10Safety, cost and the human in the loopSecurity, cost and skills; keeps general practice apart from what I measured in the earlier trials.
- 11The take-home kitStart with a short AGENTS.md, runnable checks and contracts for unread dependencies; add each further layer only when it has earned its keep.
- —Cheat sheet, glossary, references, about the evidenceBack matter in the Guide.