AI agent governance: a practical checklist
Governance for AI agents is not a policy document. It is a set of controls that still hold when the agent has been told to ignore them. This is the checklist we use.
Why agents need their own governance
Traditional access control assumes the user decides what to do. With an agent, a language model decides, and what it reads can change its mind: an email, a web page, a tool's own description. So the controls must sit outside the agent, where its instructions cannot reach.
The checklist
- Identity: every agent is a named thing with its own connection, not a shared key.
- Least privilege: each agent gets the smallest set of tools and the lowest level that does the job.
- Risk classes: every tool is classed by what it does (read, write, share, send, delete, settings, forwarding).
- Human approval: actions that leave the organisation or cannot be undone wait for a person.
- Stronger proof for the irreversible: a passkey, not a click.
- Limits: a cap on actions per day, and an automatic pause when an agent asks for too much too quickly.
- Separation of duties: people who approve are not automatically people who change the rules; some actions need two people.
- Audit: a tamper-evident record of every decision, kept apart from the agent.
- Off switch: one action that stops an agent, and one that stops an account.
- Review: look at the numbers every week and adjust.
Common mistakes
- Trusting the tool description to say how risky a tool is.
- Giving the agent the credentials. If the agent holds the key, the only control is its own judgement.
- Logging to a place the same system can rewrite.
- Approvals with no expiry, which become standing permissions.
- One all-powerful agent instead of several narrow ones.
Where Operon OS Agenaut fits
Agenaut implements this checklist as a service in front of your agents' tools. It holds the keys to the tools, so the agent never does. You set a level per agent, write rules in plain words, cap daily actions, and stop an agent in one tap. Owners and approvers are separate roles, and an account can require two people for what cannot be undone. Every decision is a signed receipt you can check and export.
Agenaut is in beta and holds no compliance certification. What it gives you is the control and the evidence.
Try it. Operon OS Agenaut is in beta. Request access, or read the docs to see how an agent is connected.
Related guides
- What is an MCP gateway, and when do you need one?
- MCP security: tool poisoning, prompt injection and how to defend an agent
- An audit trail for AI agents: what to record and how to prove it
Terms used here are explained in the glossary.