Automating the check: did the agent actually do the thing?
After three weeks trying to automate verification of agent work, a builder found their own grading disagreed with itself half the time. The post captures the core reliability problem with agents: the worker should never be its own checker, and disagreements between independent checks are the useful signal.