Most agent failures I see start as a missing stop rule. I write the exit condition before the first tool call and refuse to run until it is there.