Enterprise AI agents: approvals, roles and the two-person rule
One person approving their own agent's actions is a start. A team needs roles, cover for when someone is away, and evidence an auditor will accept.
Roles
Separate two jobs. Owners decide the rules: which agents exist, what level each has, which tools are on. Approvers answer requests: they can read everything and say yes or no, and they cannot change the rules. Most people on a team should be approvers.
The two-person rule
For actions that cannot be undone, one approval can be one mistake. A two-person rule means the first yes is recorded and the request keeps waiting until a second, different person approves. Declining takes one person.
Evidence for audit
- Who approved what, and when, with what proof.
- That the approval was for the exact request that ran.
- That the record has not been altered since.
- Changes to roles and rules, recorded the same way.
- An export an auditor can read without access to your systems.
This is the material a security or compliance review asks for. Having it does not make a product certified, and you should ask any vendor, including us, what certifications they actually hold.
Practical points
- Give each agent its own identity; do not share one connection across a team.
- Decide who approves when the usual person is away.
- Remove access the day someone leaves; their sessions should end at once.
- Keep credentials for tools out of the agents and out of chat.
How Operon OS Agenaut supports a team
An Agenaut account can have several people: owners, and approvers who can decide requests and manage only their own sign-in. An owner can require two people for what cannot be undone. Removing a person ends their sessions. Every decision and every change to roles or rules is a signed receipt in the account's ledger, which can be exported. Agenaut is in beta and holds no compliance certification today.
Try it. Operon OS Agenaut is in beta. Request access, or read the docs to see how an agent is connected.
Related guides
- Add an approval step to Claude, ChatGPT and Cursor
- Human in the loop for AI agents: approve risky actions before they run
- AI agent governance: a practical checklist
Terms used here are explained in the glossary.