AI agents are moving from demos into real work: answering tickets, updating records, running checks across systems. Each one acts with real permissions. AI agent governance is the practice of making sure every agent has a clear owner, only the access it needs, a way to have that access removed, and a record of what it did — the same disciplines identity teams already apply to people, adapted to software that decides for itself.
Start with the questions an auditor will ask
For every agent, you should be able to answer:
- What is it, and who owns it? A named, accountable person — not a team mailbox.
- What can it access? Which systems, which tools, which data.
- Who approved that access, and when was it last reviewed?
- What has it actually done? Not what it was allowed to do — what it did.
- What happens when its owner leaves?
If any answer is “we don’t know”, that is where to start.
The six practices
1. Register every agent as an identity
An agent should exist in your identity system like an employee or a service account, with a type, an owner and a status. Unregistered agents are the shadow IT of this decade.
2. Tie every agent to an accountable owner
Owners change teams and leave. Decide in advance what happens then: the agent is reassigned or suspended, never left running with nobody responsible. “Agents don’t offboard themselves” — their owners have to.
3. Grant least privilege, per tool
Grant access to specific tools on specific servers, not “all of the CRM”. Use short-lived, scoped credentials where you can.
4. Review agent access like anyone else’s
Put agents into your access reviews. Reviewers need context — when the access was last used, how it compares with similar agents — or reviews become rubber stamps.
5. Enforce at runtime, not only on paper
Permissions describe what an agent may do; enforcement decides what it can do. Put a checkpoint on the path of every tool call so that scope, rate and data rules are applied as the agent runs (see What is an MCP gateway?).
6. Keep one audit trail for people and agents
When an agent acts for a person, both should appear in the record. A tamper-evident trail means the record can be trusted when it matters — in an incident or an audit.
How AtmanID approaches it
AtmanID governs humans and AI agents on one identity model and one tamper-evident audit trail (What is AtmanID). Agents are registered identities with owners; access is certified in campaigns with reviewer context such as last use and peer comparison (launch a certification campaign); and every agent → MCP tool call passes through the AtmanID gateway for scope, rate and data checks (MCP gateway). Separation of duties covers agents too: rules name MCP tools one agent must not hold together, or a tool an agent must not hold while the person it acts for holds a conflicting entitlement, and findings include two agents of one person splitting a task and agents that can call each other (Separation of duties for AI agents). Agents’ tool grants are certified in Agent Access Reviews, where the agent’s owner sees when each tool was last used.
Related: AI agent security: where identity fits · What is a non-human identity? · What is an MCP gateway?