If you’re letting an AI agent touch production, a green unit test is not enough. The failure mode is rarely “the model was dumb” — it’s usually missing guardrails around scope, rollback, credentials, or observability.
Here’s the lightweight checklist I’d want before any agent can run outside a sandbox:
1) Make the blast radius obvious
What files, APIs, repos, or environments can it touch?
What is strictly read-only?






