“I spent, like, half a day trying to convince it not to do it. It was bent and determined to model it the way it wanted to. So I got one of my agents to debug the other agent. I got Claudio to go figure out what Ed was doing. And Claudio comes back and says: oh, well, yeah, of course it’s gonna do that — because in the Ed agent, you’ve explicitly said…” — Lou
Session context: 2026-07-23_Mastermind — Lou was trying to teach his editor agent to model the structure of a Michael Simmons article rather than its content, and the agent kept copying the content instead. Half a day of arguing with it, and the cause turned out to be a rule Lou had written himself and forgotten.
Core Idea
The agent wasn’t disobeying. It was obeying something you no longer remember writing.
Lou asked his editor agent to extract the meta-structure of an article — the skeleton, the flow, the thought process — explicitly not the topic. What came back was a literal transplant: not “insert a citation here” but “insert a citation from Charlie Munger,” because Munger was cited in the source. Repeated correction did nothing. The model was described as more literal in following instructions than earlier generations, and it was being literal about an instruction Lou had given it months earlier, buried in a rule file, that forced explicit execution over meta-level abstraction.
This is a specific and under-named failure mode. Prompt-level correction cannot beat a rule-level constraint. You can rephrase your request twenty times and lose the entire afternoon, because you are arguing at a layer that gets overridden before your words are even considered. The more disciplined you have been about writing rules — and the more the model has improved at following them — the more likely this becomes. Instruction-following is a capability that cuts both ways: an agent that reliably obeys reliably obeys your mistakes.
The move that broke the deadlock is the transferable part. Lou stopped debugging the output and had a second agent debug the agent. Claudio read Ed’s instruction set as source code, found the conflicting rule, and explained the behaviour in one pass. That reframe matters more than the fix: an agent’s instruction set is a codebase, and when behaviour is inexplicable you read the code, you don’t argue with the runtime.
Scott sharpened the structural requirement in chat, and it belongs in every implementation of this: “the auditor agent can only report to the orchestrating agent, not any agent it’s auditing.” An auditor that talks to its subject is a negotiation, not an audit.
The scaling problem Lou named next is real. At a handful of agents you can read every instruction file yourself. Past that you cannot, especially if any of them are self-improving. The audit function has to become a standing part of the system rather than something you do when you get frustrated — and it’s the argument for a canonical library, where every skill lives in one place so a single pass can find the contradictions and one fix propagates everywhere it’s used.
Practical Application
Next time an AI ignores an instruction twice, stop rephrasing. Open a fresh session with a different agent and give it this: “Read every instruction file governing agent X. I asked it to do A and it did B instead. Find the rule that produced B. Quote it and explain the conflict.” Budget ten minutes. Compare that to the half-day Lou lost. Then run the same pass across your whole library quarterly, asking specifically for rules that contradict each other between agents doing similar work — the two-writer-agents-with-opposite-editorial-guidelines case Lou described is the common one.
Related Insights
- Insight - Skill Slop — Unconstrained Skills Get Invoked in Unintended Ways — skills firing when they shouldn’t; this is the inverse, a rule firing correctly and doing damage.
- Insight - The Universal Audit Phrase — Errors, Omissions, Oversights, Duplications, Contradictions — the audit vocabulary to run against an instruction library, not just a plan.
- Insight - The Self-Improving Skill Loop — Have the Skill Learn From Every Use — self-improvement is what makes an unaudited instruction set drift beyond what you can hold in your head.
- Insight - Your Rubric Can Certify Garbage — When the Evaluator Drifts With the Work — the same session’s sibling failure, one layer up: the evaluator drifted rather than the executor.
- Insight - Fewer, Better Agents — And the Auditor Never Reports to the Audited — the structural constraint that makes agent-audits-agent trustworthy.
Evolution Across Sessions
Builds on Insight - Skill Slop — Unconstrained Skills Get Invoked in Unintended Ways (2026-02-05), which established that under-constrained skills get invoked in unintended ways. The new development is the opposite pathology: over-constrained agents, obeying rules their author has forgotten, producing behaviour that looks like disobedience and is actually perfect compliance. It also introduces the diagnostic move — agent debugs agent, reading instructions as source — and Scott’s structural rule that the auditor must report upward to the orchestrator rather than to its subject. Together with Insight - Your Rubric Can Certify Garbage — When the Evaluator Drifts With the Work, this session establishes a theme worth tracking: as the agent layer thickens, most of your bugs will be things you wrote and stopped remembering.