# AI Agent Governance Review

A short, honest checklist before putting an agent in production. Based on the
AgentAz Trust Levels — https://www.agent-kits.com/agentaz-specification

## 1. Worst-case action
- [ ] We can state, in one sentence, the worst irreversible thing this agent could do.
- [ ] That action is either impossible, or gated behind human approval.

## 2. Tool boundary
- [ ] Every tool is sorted: read-only, reversible write, or irreversible/consequential.
- [ ] Every irreversible/consequential tool requires human approval before it runs.
- [ ] The approval gate lives outside the model (it can't be prompted away).

## 3. Trust Level
- [ ] The agent's declared Trust Level (ADV / A0 / A3 / A4 / A5) matches its tool boundary.
- [ ] A change that raises the Trust Level triggers a re-review.

## 4. Bounds
- [ ] Cost, loop, and rate limits are enforced (not just requested in the prompt).
- [ ] There is a clear stop condition / kill switch.

## 5. Ongoing
- [ ] agentaz.json is committed and validated in CI (agentaz-validate GitHub Action).
- [ ] Governance drift is monitored as the agent changes over time.