With the control owners established, identify who will carry out the work and what each identity can reach. Include the person who starts an agent, the service account that executes it, and any credentials passed to its tools.
| Standard example control | Implementation steps |
|---|---|
| Agent inventory | Inventory enabled agents and flows, with a named human owner. |
| Agent identity | Trace the initiating user and executing identity. |
| Secrets management | Limit credential scope and lifetime, then use the revocation procedure when access ends. |
Inventory enabled agents and flows
Open the group's and project's AI > Agents and AI > Flows pages. Record which items are enabled, the workflows that use them and the person responsible for each. Use the AI Catalog to inspect their definitions and managing projects. For custom items, record the selected version and who can update it; GitLab-managed foundational agents and flows don't use the same version-pinning model. Include integrations and service accounts that aren't represented in the catalog in the same inventory.
Reconcile service accounts with the inventory
Set GSS_GROUP_ID to the numeric ID of the group being reviewed, then list the group's service accounts:
glab api --hostname gitlab.com --paginate "groups/${GSS_GROUP_ID}/service_accounts?per_page=100" | jq --slurp '[.[][] | {id, username}]'
Check that every service account appears in the inventory with an owner. Reconcile the inventory against AI > Agents and AI > Flows at group and project level. Some enabled toggles and external integrations have no single readable API.
Trace the initiating user and executing identity
For runner-executed DAP flows, GitLab's composite identity limits access to the intersection of the initiating user's and service account's permissions. Inspect both memberships, including other projects they can both access. Record the flow session and resulting commit or merge request so another operator can identify the human initiator and the automated action. Agentic Chat in the UI and IDE doesn't use this same composite-identity mechanism, so inspect its authorization separately.
A service account's involvement doesn't make a change independently reviewed. Keep the initiating author separate from the human who approves the resulting change.
Read the merge request author and commit
Set GSS_APP_ID to the application's numeric project ID and GSS_MR_IID to the agent-authored merge request's number within that project. Read the author and commit:
glab api --hostname gitlab.com "projects/${GSS_APP_ID}/merge_requests/${GSS_MR_IID}" | jq '{author: .author.username, sha, web_url}'
Link one flow session to its initiating user, service account and resulting commit or merge request. Check that the executing identity is recorded and differs from the approver. The link between session and human isn't a single API field.
Limit credential scope and lifetime
For each credential the workflow uses, record its issuer, owner, purpose, reachable systems, expiry and revocation procedure without recording its value. Give a task only the access its next action needs. Keep release and policy-administration credentials outside application-controlled jobs. Use the organization's approved secret store; GitLab documents Secrets Manager and external secret providers as retrieval options. Choose the integration for the actual executor and review who can request each secret before supplying it to a flow.
Choose the workflow's delegated authority
With ownership and permissions defined, decide how much work an agent may carry out. The Standard makes that decision per workflow:
- Suggest requires a proven Foundation and Authorize stage. The agent recommends; a person makes the change.
- Propose requires the Foundation and stages through Verify. The agent can open a merge request; a human approves release.
- Act with guardrails requires the Foundation and all five stages, including a tested stop, with limits on volume and blast radius.
Apply these levels to individual workflows. A team can use Propose for application repairs while keeping changes to security controls in Suggest. Neither decision grants the same authority to another workflow, and either can be revisited when its controls change.
Write the authority decision so another practitioner can understand it. Name the allowed task and resources, prohibited actions, approval points, stop conditions, credential issuer and lifetime, records to retain, and accountable owner. For Act with guardrails, define enforceable limits on both the amount of work and the systems it can affect: for example, changes per run and the permitted projects or environments. Avoid blanket statements such as “the agent can maintain the project.” They make it difficult to distinguish intended authority from an accidental capability.