Field Notes

Patterns for shipping AI agents where mistakes are expensive. Each note explains the problem, how the pattern works, where it breaks, and lets you step through a scripted run.

The runs are illustrative, with made-up companies and numbers. The patterns are the ones I use.

01Human always sendsAgents draft and fill forms. A person owns every message that leaves the building.3 interactive runs 02Confidence gatingLet the model act only where it has earned it. Everything else gets a head start, not an autopilot.3 interactive runs 03Evals before scaleDecide what "good enough" means, measure it on real cases, and write the release rule down before the results come in.2 interactive runs 04Staged release gatesMatch the controls to the risk. Let low-risk changes flow, and make high-risk ones earn their way out in stages.3 interactive runs 05Orchestrators, supervisors, and MCPSplit planning from doing, give every agent the narrowest tools that work, and have something watching what they actually try.3 interactive runs