The Agentic AI Playbook

1. Which decisions an agent should be allowed to makeStart with the decision, not the model. Most agent projects fail because nobody wrote down what the agent is allowed to do.
2. Guardrails that live outside the modelA rule inside the prompt is a suggestion. A rule in code is a guardrail. Know the difference and build both.
3. Step limits, so a wrong call stays smallThe most useful guardrail is boring: no single decision may move anything by more than a bounded amount.
4. Evaluation before anything shipsIf you cannot score a change against a fixed set of real cases, you do not know what you shipped.
5. Cutting AI cost 10x with a model cascadeMost of the work an agent does is sorting, not thinking. Do not pay frontier prices for sorting.
6. Reversal paths and audit trailsEvery action an agent takes should be undoable for a window, and explainable forever.
7. Real incidents, and what we changedThree things that went wrong running agents in production, and the fix for each. Names and numbers simplified.

Who this is for

Founders deciding whether to let AI act in their business. CTOs and engineering leads building their first production agent. Product and operations people who will own the decisions an agent makes.