Yeah, they've got it backwards. I tried to sum it up in thisistheway.to/ai but what's been working for us is that every agent miss is a learning opportunity:
1. Capture the miss — What did the agent do? What did reality say?
2. Diagnose — What didn't it see? Missing data, constraint, feedback, or boundaries?
3. Choose a primitive — Observability, instructions, tooling, guardrails, or verification?
4. Encode as artifact — Version-controlled, repeatable, not just memory.
5. Promote to gate — When it's worth enforcing, make it a gate.
Every harness I setup includes this process in the primary set of agent instructions.
1. Capture the miss — What did the agent do? What did reality say?
2. Diagnose — What didn't it see? Missing data, constraint, feedback, or boundaries?
3. Choose a primitive — Observability, instructions, tooling, guardrails, or verification?
4. Encode as artifact — Version-controlled, repeatable, not just memory.
5. Promote to gate — When it's worth enforcing, make it a gate.
Every harness I setup includes this process in the primary set of agent instructions.