This article is general information, not an audit opinion or compliance advice. Agree system scope, applicable criteria and evidence expectations with your auditor.
Quick answer: Include agents when their access and activities are relevant to the system and commitments covered by your SOC 2 engagement. Document the controls and evidence for provisioning, permissions, monitoring, incident response and removal of access. A named identity and retrieval history can support that work, but neither guarantees a particular auditor conclusion nor proves every downstream action.
Start with the engagement, not the agent label
SOC 2 evaluates controls against applicable Trust Services Criteria within a defined system and engagement. The AICPA criteria resource is the primary reference; use the applicable text with your auditor.
An agent may generate code, administer infrastructure or access customer records. Identify which activities and supporting services are relevant to your system description and commitments. “It is AI” neither establishes a finding nor makes a workload exempt.
This guide groups practical questions under CC6, CC7 and CC9. It does not claim to validate each detailed subcriterion or prescribe sufficient evidence for a particular engagement.
CC6: explain access and its lifecycle
For each relevant agent, be able to explain:
- Its owner, task, environment and runtime.
- Who approved access and how it was provisioned.
- How the caller authenticates and which credentials it can retrieve.
- What the underlying provider credentials actually permit.
- How access changes and removal are authorised and verified.
Separate identities can make retrieval attribution and offboarding easier. Shared credentials can complicate those activities, but a shared key alone is not an automatic audit conclusion. Assess the full control design and evidence.
Do not equate a named profile with least privilege at the provider. A retrieval policy allowing a broad key does not narrow that key once delivered.
Keep records of approvals and changes, not just a screenshot of today's configuration. Review access at a documented cadence appropriate to risk and the engagement rather than presenting one interval as universally required.
CC7: show monitoring and response in operation
Useful evidence can include monitored events, investigation records, escalation decisions and completed containment exercises. An installed logging product is not the same as a control demonstrably operating.
Separate three evidence layers:
| Layer | Evidence it can contribute | What not to infer |
|---|---|---|
| Credential retrieval | Caller, requested credential, time and release outcome | Every provider operation afterward |
| Runtime or tool execution | Tasks, tool calls and execution results, depending on instrumentation | Complete coverage or absence of tampering |
| Provider activity | Recorded operations, resources and provider identity | Agent-level attribution when keys are shared |
Correlate records where supported and document gaps. Protect logs from inappropriate access or alteration, and exclude secret values and unnecessary customer data.
Response procedures should address both new key delivery and credentials already held. Revoking a retrieval profile does not automatically stop direct API calls from a runtime with a retained key. Where compromise is suspected, stop or isolate execution where possible and invalidate exposed provider credentials as appropriate.
See the incident-response playbook for those separate actions.
CC9: review dependencies and vendor risk
Identify the model service, hosting platform, agent client, credential store, monitoring systems and other relevant third parties. Review their actual roles, data access, security responsibilities and incident commitments.
A component that receives only retrieval metadata has a different data flow from one receiving customer records. Neither description should be assumed without examining the integration.
Document onboarding, changes and termination of these relationships. A vendor's features or report do not replace assessment of your own configuration and responsibilities.
Additional approval is useful only where enforced
For sensitive exports, deployments or other high-risk operations, approval or deterministic checks may support your control objectives.
The execution system must enforce the decision on the actual operation and parameters. An agent with another usable direct key may bypass an approval-only interface.
Keep the approving identity, decision and execution evidence where relevant. Do not promise that an approval gate is a universally accepted compensating control or guarantees a favourable auditor response. Its value depends on design, operation and the engagement.
Prepare an evidence set with owners
A practical starting set includes:
- The system description and relevant agent/data-flow inventory.
- Provisioning, permissions, approval and access-change records.
- Access-review decisions and evidence that removals were verified.
- Retrieval, runtime and provider evidence with documented coverage.
- Incident exercises, investigations and remediation records.
- Third-party review and responsibility documentation.
- Retention, access and integrity controls for evidence.
A Type II report evaluates operating effectiveness over its specified period. Agree that period and evidence expectations with the auditor; do not assume every engagement or raw log requires 12 months. Implement controls early enough to produce the evidence needed for the period being examined.
Where Gazebo can contribute
Gazebo provides named retrieval identities, selected-key approval for OAuth connections, retrieval policies and credential-access history. These can make credential access easier to manage across agents and services.
It returns the actual stored provider key to the runtime and its host. Provider permissions govern subsequent calls; Gazebo does not approve or log those operations. Revoking retrieval does not invalidate retained provider keys.
Use those capabilities as evidence-producing components within your control environment, not as a SOC 2 compliance claim. For operational ownership and reviews, see the credential-management lifecycle.