What they did
Shawn Knight was working on A.I.N.D.Y., his own execution runtime, and wanted to know whether its Memory Bridge component actually did what he believed it did. He first used Claude in chat to plan audit prompts, carried them to Claude Code, and brought the findings back by hand. Claude Code searched the repository, compared several half-finished memory implementations against each other, checked which one was really wired in, and ran the tests. Before changing anything it recorded a baseline, then wrote tests for the behaviours the audit found missing. The rebuild ran in three phases (make it real, make it intelligent, make it useful), with the agent told to read named files first, keep stated architectural constraints, run tests after each phase, and stop to investigate when the code contradicted an assumption.
How it went
The baseline was 135 passing tests plus 8 that failed on purpose, one for each missing behaviour. Over the three phases the suite grew to 338, then 362, then 384 passing with none failing. Knight says the audit also changed his own idea of the component: it was meant to orchestrate recall and feedback, not just store data. Over time he stopped relaying work through the chat and did the planning inside Claude Code, because it could look at the real files. He admits agents still got things wrong and later agents caught misses by earlier ones.
Worth knowing
The test counts are the author's own report and he says plainly they do not prove the system is correct. The useful trick is writing the missing behaviour as deliberately failing tests before the rebuild, so progress is measured by the repo and not by the agent's own summary. The post is part of a series and the source code is linked from it.