Separation of duties for AI agents
Separation of duties (SoD) says no one should be able to start a sensitive task and also approve it. For people,
AtmanID checks pairs of entitlements (Configure SoD rules). AI agents do their work
through MCP tools, so agent rules are pairs of tools: an agent that can both create_payment and
approve_payment can pay itself.
Agent rules are on Security & Risk → SoD & Policy Engine, on the Agents tab. Identity Admins, PAM Admins, Tenant Owners and Super Admins see the Add Suggested Rules and New Rule buttons.
What a rule covers
An agent rule names one MCP tool and what it must not be combined with:
| Must Not Be Combined With | The conflict |
|---|---|
| Another tool (same agent) | One agent’s grants on that server include both tools. |
| An entitlement the person it acts for holds | The agent holds the tool, and the person it acts for holds the entitlement — together they hold both halves. |
A rule belongs to one MCP Server: AtmanID’s own server, or any server you registered. A tool name only means something on its own server, so the same names on another server are not a conflict.
An agent “holds” a tool when one of its active, unexpired grants on that server lists it. A grant with no tool list allows every tool the server offers, so it holds every tool a rule names there.
Add rules
- On the Agents tab, click Add Suggested Rules for AtmanID’s own tools — for example Request and approve access, Grant access and certify it, Create a role and assign it. They are added switched off; switch on the ones that apply.
- Or click New Rule and fill in New Agent SoD Rule: Rule Name, Description, MCP Server, Tool, Must Not Be Combined With (a Second Tool or an Entitlement), When a Grant Gives Both, and Risk Level.
Switch On and Switch Off on a rule’s row turn it on and off.
When AtmanID checks
- When a grant is made. Granting an agent tools on a server is checked against your active rules. The grant goes through and the conflict is recorded and shown under Open Conflicts. Rules warn for now: a Block rule refuses a single-agent grant only once enforcement is switched on for your tenant.
- Every day. Every active grant is checked again, and conflicts that no longer hold — the grant was revoked or narrowed — are resolved.
- When access changes around the agent. Granting or revoking agent-to-agent access, changing an agent’s owner, and delegation changes re-check the findings that involve other agents and people.
Findings across agents and people
One agent with both tools is the simple case. Across Agents and People lists the ways a rule can be split:
| Finding | What holds both halves |
|---|---|
| Agent and person | The agent holds the tool; the person it acts for (its owner, or the human at the root of a delegation chain it is in) holds the entitlement. |
| Two agents, one person | One person owns or delegates to two agents, and each holds one tool of the pair. |
| Call chain | An agent holds one tool and can call another agent (up to three hops, within each link’s allowed capabilities) that holds the other. |
| Same record | Two different agents used the two tools on the same record — one created the payment, another approved it. |
Same record needs the record’s key: AtmanID reads it from tool calls on registered MCP servers that store call payloads, and keeps only a hash of it. AtmanID’s own server keeps no payload copy, so same-record findings come from your other servers. When the two agents share an owner or can call each other, the finding says so; unrelated agents are still shown, so a reviewer can decide whether it matters.
Findings are warnings. They resolve on their own once the access that caused them is gone.
Certify agents’ tool access
An Agent Access Review puts agents’ tool grants in a certification campaign, next to people’s access:
- On Certifications, launch a campaign and choose Agent Access Review. Leave the scope empty to review every agent tool grant, or enter one MCP server.
- Each active grant becomes one review, assigned to the agent’s owner. If your certification settings forbid self-certification, an owner certifying their own agent’s access counts as self-certification, so the review goes to the owner’s manager instead. Grants with no one to review them (no active owner, or no active manager) are left out, and the launch message says how many.
- The reviewer sees the agent, its tools (“All tools on this server” for a grant with no tool list), the server, and when the agent last called each tool through the MCP gateway — so tools it never uses stand out.
Revoke removes the grant straight away: the grant, the tokens issued under it and the findings that depended on it. Revoke agent reviews one at a time; bulk revoke and Iris do not revoke them. Certify keeps the access. Every decision is in the audit trail.
Related
- Configure SoD rules — rules between entitlements, for people.
- MCP gateway policy engine — what is checked on every agent tool call.
- Launch a certification campaign.
© AtmanID · Privacy