How do you govern AI agents at work: what they may do, who signs off, and how you prove it?
This is not about building agents but about running them inside a company. How do teams decide what an agent is allowed to do, who has to approve its actions, and how they show afterwards that the work was done correctly?
Treat an agent like a new contractor with a narrow badge. Write down which tools and systems it may touch, with what access, and default everything else to no. Read-only access covers a surprising amount of useful work, and each write permission should be added on purpose with a named owner.
Sign-off works best when it is tied to risk, not to every step. Let the agent act alone on reversible, low-cost actions, and require a person to approve anything that sends money, deletes data, contacts customers or changes production. Put that rule in the tooling as a policy, not just in the prompt, because a prompt can be talked around and a policy cannot.
Proof means a record you did not have to ask the agent for: which tool was called, with what input, what came back, who approved it and when. Keep it outside the agent's own control so it cannot be edited after the fact. When something goes wrong, that log is what lets you explain the outcome instead of guessing.
Finally, review a sample of completed work regularly, the same way you would check a new hire. A separate review pass against your security and quality rules catches drift that no single approval would notice.
Listings mentioned
- protect-mcp — Policy Enforcement + Signed Receipts · skill by wshobsonSets up policy-gated tool calls and signed receipts for Claude Code, giving an audit trail of what the agent did.
- Security reviewer · agent by shinprReviews finished work against your security requirements and returns only the must-fix findings.
Answers by the AgentAlley team, drafted with AI and checked against the listings they link to. Not a real-person reply from the original thread.