Skip to content
Index / writing

Writing on agent reliability and LLM cost

How I think about making agents trustworthy and affordable in production — the same methods behind the audits and teardowns.

False Green

1-2 emails a month

measured notes from running AI agent systems in production: what broke, what it cost, what the data showed. 1-2 emails a month.

October 2026

5 posts
engineeringHow to audit your own Claude Code and Codex session logs, locally, in one commandengineeringA Claude Code Stop hook that checks a test actually ran after the last editengineeringDoes Codex CLI have hooks? What they send, and what they leave outarchitectureWhat is a false green in an AI coding agent's work?architectureI built a tool to catch agents saying "tests pass" without proof. It was wrong 43 times out of 46.

September 2026

5 posts
engineeringThe Check the Agent Cannot Skip: 0 of 40 Silent Failures vs 22 of 402260Silent failures / 40 runsengineeringSelf-Verification Cost 1.64x and Caught Nothing1.64xcost · 0 catchesengineeringThe API Served a Different Model and My Check Passedrequesteddeepseek-v4-flashserveddeepseek-flash200 / 200 responses · requested ≠ servedarchitectureNine Times the Eval Tried to Lie to Me9times the eval liedarchitectureWhen an AI Agent Says Done: 0% vs 100% False Greens0%vs100%false greens · one answer vs trajectory

June 2026

5 posts
architectureThe Arc Since the Six Were CaughtarchitectureThe Calibration Ledger: 58 Runs, 93% Pass, n Is Small~93%~58 runsn is smallarchitectureFalse Greens: Three Structural ObservationsarchitectureOne-Sided Contracts Break Agent PipelinesarchitectureQA Is Only as Honest as Its Coverage

May 2026

2 posts
architecturePermissive Parsing Is a False-Green FactoryarchitectureSix False-Greens in a Self-Auditing Agent Pipeline6false greens caught