---
url: /handbook/governance/request-a-governed-agent.md
description: >-
  From a request to an issued agent, with each decision recorded, and what issue
  does and does not do.
---

# Request a governed agent

This page follows one request from start to finish.

### The employee asks

An employee asks for an agent. The request can start in three places:

* in DutyGraph, from the agent builder;
* from the employee's own AI chat, over MCP (see [Request from your AI](request-from-your-ai.md));
* from an opportunity a manager assigned to them ("Suggested by `<manager>`"). It becomes a request only when the employee submits it, so the requester is always the person who asked.

The request names one or more Task Cards. Each must exist in the governance registry.

### DutyGraph compiles the proposal

If the requester names no scopes, DutyGraph starts from the scopes recorded for the chosen Task Cards. The requester can also name scopes.

The evaluator then checks each scope against the System Catalog, the owner's holdings and the rule packs. Each scope gets one state: **Inherited**, **Auto-trimmed** or **Escalated**, with a reason and a rule citation.

The evaluator does not check named scopes against what the Task Cards need. That rule, CEIL-02, is not enforced yet. See [The key join](the-key-join.md).

This step is code. No AI call is on this path. The requester may answer at most one clarifying question, offered as a choice, for example "Atlanta only / all branches".

### SoD runs on three sets

The evaluator checks the owner alone, the agent's scopes alone, and the owner plus all the owner's active agents as one set. Conflicts are shown with the rule text, one click from the scope.

### The approval chain decides

Manager, role owner, Internal Controls and IAM each choose **Approve as proposed**, **Approve with changes** (narrow only) or **Deny** (with a reason). Reviewers judge whether each scope is needed for the work. Internal Controls can keep an escalated scope only by tagging a mitigating control that covers agents, usually with a time limit.

### IAM issues the agent in DutyGraph

Issue writes a ledger entry, sets the agent's end date and creates one credential for the DutyGraph gateway. It returns a receipt that lists each trim and refusal with its rule.

A "For IAM to do" card names the agent identity to create, with exactly the approved scopes. IAM does that in the company's own systems. DutyGraph does not provision.

Before issue, DutyGraph checks a fixed table of runtime controls, such as stopping the agent at its end date. If the chosen runtime is marked unable to hold a mandatory control, issue to it is refused. The table lists "Ask before supervised calls" as supported for every runtime, but the DutyGraph gateway approval gate it names is not built. See [Agent gateway](agent-gateway.md).

### Calls through the gateway are checked

If the agent calls its tools through the [agent gateway](agent-gateway.md), each call is allowed or blocked, and each block cites its rule. Calls the agent makes some other way are not checked or recorded by DutyGraph.

## Approval, provisioning and enforcement are different

| Step | Who | What it shows |
| --- | --- | --- |
| Approval | The approval chain, in DutyGraph | People decided, and the ledger recorded it |
| Provisioning | The company's IAM | The identity and access exist in business systems. Not recorded in DutyGraph today. |
| Enforcement | The gateway, the runtime and the business systems | A call was allowed or refused. DutyGraph sees only gateway calls. |

See [Approval, provisioning and revalidation](approval-provisioning-and-revalidation.md).

## After issue

* IAM or Internal Controls can revoke the agent. Its credential stops working at once.
* The gateway blocks calls after the end date.
* If the owner leaves, their agents are **not** suspended automatically. Revoke them by hand.

::: info
**Status.** **Built**: every step above, the "For IAM to do" card, revocation and expiry. **Partial**: task-based justification of each scope (CEIL-02 is not enforced). **Simulated**: in the sample companies, the identity systems, the business systems and the agent runtime, and the screens say so. **Planned**: CEIL-02, provisioning in the company's IAM, Decision Export and owner-departure suspension.
:::
