Skip to content
Checklists/10 checks

AI Agent Safety Checklist

Before an agent runs unattended, every answer here should be yes. A single no is a reason to keep a human in the loop.

0 / 10
Worked review · maintained by TianMingAI · 2026-07-21

A completed example

An agent is asked to update one documentation generator. It works in a disposable worktree with access only to the repository, while production credentials, external messaging, unrelated directories, and automatic deployment remain out of bounds.

Pass

The write scope is narrow and isolated

The task permits changes only to the generator, its tests, and the matching documentation page inside one worktree.

Pass

The agent cannot use production credentials

No production secret is available in the worktree and the task requires no remote write.

Stop

Unrelated edits are rejected

The draft also changes a shared navigation component, so the checker stops the run before approval.

Pass

A rollback path exists

The unapproved worktree can be discarded without changing the stable branch.

Decision

Do not approve the draft while an unrelated navigation change remains. Isolation contains the risk, but it does not make the out-of-scope edit acceptable.

Next action

Remove the unrelated change, rerun the generator tests and full checks, then ask the independent reviewer to compare the reduced diff with the original request.

Why this checklist matters

Agent safety is mostly determined by the surrounding permissions, checks, and approval rules. A capable model operating with vague limits can still make an irreversible mistake faster than a person notices it.

How to use this checklist

Review the checklist with the real credentials and tools the agent will receive.
Test a deliberate failure, repeated error, and out-of-scope request before unattended use.
Treat every new permission or external action as a reason to repeat the review.

Decision after the check

A single missing control around irreversible actions is enough to keep the run supervised. Low-risk gaps may be accepted only when a named person is monitoring and can stop the run.

Review questions

What is the worst action these credentials allow?
Can the agent hide a shortcut behind a passing result?
How quickly would a person learn that the loop is stuck or unsafe?

Related templates