The Config Discipline Rule: Why Your AI Agent Problems Are Probably Not Model Problems

Most AI agent debugging sessions end the same way. You spend 45 minutes tweaking prompts, adjusting temperature, trying a different model. Then you look at the SOUL.md and realize: the identity file has not been updated in three weeks. The task spec is two sentences. There is no exit condition. The agent was not broken. The config was. The Config Discipline Stack Reliable agents are not smarter than unreliable ones. They have better config discipline. 1. Identity file (SOUL.md) — current and specific Every behavior problem traces back to identity drift. If SOUL.md has not been reviewed in 30 days, assume it is partially wrong. Review it before debugging anything else. 2. Task spec — three fields minimum scope : what the agent touches and what is off-limits done_when : the observable condition that means finished success_criteria : what good output actually looks like Without these, the agent guesses. The model does not cause the problem — the missing spec does. 3. Exit condition — expl