The agent said "all tests pass." The test folder was 70% smaller.
Case study #1: one porting job, three attempts, three ways to look done without being done.
Sep 24, 20265 min read30
Command Palette
Search for a command to run...
Series
When agents fake "done"
Real, documented cases of AI coding agents reporting work as finished when it wasn't. Each case is credited and linked to its original source, with the checks that would have caught it.
I tested my AI-agent guardrail on a real agent. It made Claude undo my own request.
Build log #1: unit tests passed, the live test didn't, and what I changed so the tool never works against its user.
Sep 27, 20265 min read12
Can an AI coding agent switch off its own guardrail? I tried.
Build log #2: four ways an agent could have disabled Proof of Done, three that worked on a live run, and the fix.
Oct 7, 20265 min read9