About the evidence

Free to read on the web. October 2026 PDF · Updates

My numbers. Every figure in a case study comes from the repositories and reports above, with the file named next to it. Every chart was generated by a script from tables in those files. Samples are small: most cells are 3 to 16 runs, many external-validity cells are single runs, and most experiments used one agent (OpenAI’s Codex, recorded as gpt-5.6-luna at medium reasoning). The worked-example agent trials used Cursor’s composer-2.5. Several reports describe versions of the same experiment or tool, so when a number appears in two places, count it once. Read the numbers as signals from specific experiments and nothing more general.

The worked example. A reproducible engineering demonstration of the failure modes found in the agent experiments (guessed facts, stale knowledge, unverifiable claims), published at meabs/ProgrammingTheLoop-WorkedExample: a storefront-api consumer repository, an inventory-service provider, the stage files, and reproduce.sh, which rebuilds the git history from scratch, runs every stage and then checks each result against what this book reports, failing on any mismatch. How it was built is set out in Chapter 1 and the repository README; agent trials on its task are reported in Chapters 3 and 5. The “guessed names” variant is a deliberate illustration, the friction note was filed with the real tool, and the status-line format was written for this book. All test output, curl responses, the diff summary, the freshness-gate output and the friction record shown in the book are genuine output of its scripts (Python 3.13.5 with the exact package versions pinned in its requirements.txt, and AgentFrictionLoop v0.3.0 from its public repository). Timings in the quoted test output are from the run that produced the current logs/ and vary by a few hundredths of a second between runs. The final improvement in Chapter 6 (moving the live check into the repository) is described, not run.

Quotes. Every quotation was copied from the linked source and checked word for word by script before this edition was built. Quotes keep their authors’ original spelling. The keynote transcript includes filler words and some transcription errors.

Licences of my repositories (as published in each repository’s LICENSE file):

• OKFExperiment, AgentTrialRig, AgentGraphWorkTree and AgentFrictionLoop: MIT licence (“Copyright (c) 2026 meabs”).

Who ran the experiments. The experiments, tooling and analysis in those repositories are entirely my own work. Where the repository write-ups use the editorial “we” or “the experimenters”, that’s just technical-report convention; there was only one person putting the kettle on.

Who wrote the words. I used AI tools to help draft and revise the prose, the way the book recommends using them for code. Every claim, number and recommendation is mine, and I checked them against the sources.

Self-interest. Several of the practitioners discussed build or promote agent tools or techniques. I built the tools I evaluate in Chapters 3, 5, 6 and 9, which is a reason to read those chapters’ caveats carefully.