Quick answer: Before putting an AI agent into production, verify six things: runtime trust, provider permissions, retrieval identity, controls on high-risk operations, usable evidence and a tested containment path. An access profile is not a substitute for provider restrictions or safe execution.
This is a production-readiness checklist. For checks specifically about secret storage and retrieval, use the six secrets-management practices. For the underlying components, read the architecture guide.
1. Know which runtime and host you are trusting
Ready when: You know where the agent executes, which tools and files it can access, which third parties receive tool output and how credentials may be retained.
Review execution permissions, data access, logging and the handling of untrusted content. Local execution, hosted execution and background jobs can have different trust boundaries; the client name alone does not establish them.
Gazebo credential retrieval returns the actual stored provider key to the runtime and its host. Whether it enters model context, traces or conversation history depends on the client. Do not assume MCP keeps a returned key outside the conversation.
2. Verify the provider's actual permissions
Ready when: The provider credential is restricted to the task's necessary operations, resources and environment, where supported, and representative prohibited operations are rejected.
Keep production and test credentials separate. Read-only is not automatically low risk: exporting customer data or private source code may be sensitive even without a write.
Gazebo's retrieval policy decides which stored credentials an agent may receive. A declared GET method does not make a write-capable provider key read-only. If the provider cannot express a required restriction, change the integration or evaluate an architecture that controls actual requests.
Do not postpone minimum permissions until after launch.
3. Separate independently managed retrieval identities
Ready when: Agents with different owners, environments or offboarding needs can be identified and have future retrieval disabled independently.
Use named identities and approve only the required stored keys for OAuth connections. Protect the bootstrap token or connection itself; it is another access credential.
Confirm that a permitted test retrieval succeeds and an unrelated credential request is denied. Separate retrieval identities do not automatically mean separate provider keys. Shared provider keys can still complicate attribution and invalidation.
4. Control high-risk operations where they execute
Ready when: Operations requiring approval, limits or extra validation are controlled by the system that actually executes them.
Examples include charging money, changing infrastructure, deleting data and exporting sensitive records. Review the actual destination, parameters and resources—not just a natural-language description of the intended task.
An approval step is useful only if the agent cannot bypass it through a direct provider key or another tool. Validate that boundary before relying on it.
Gazebo OAuth consent approves credential retrieval. It does not inspect or approve subsequent provider operations. Operation-level approvals belong in an appropriately designed runtime, tool or proxy; they are not supplied merely by connecting Gazebo.
5. Verify evidence before relying on it
Ready when: You can find the expected retrieval records and the provider/runtime records needed for your incident questions, with secret values excluded.
Test one successful retrieval and one authenticated policy denial with a non-production credential. Check actual event coverage, access to logs and retention.
Gazebo records credential retrieval, not every provider call. One released key can be reused many times. A record of key delivery may help an investigation, but does not prove which customer record was read or which deployment was made.
6. Exercise containment and assign an owner
Ready when: Someone can block retrieval, stop or isolate the runtime where possible, and invalidate a potentially exposed provider credential.
Test these as separate actions. Confirm rejection of a new retrieval after revocation and verify provider invalidation when required. Know which other workloads share the key and how replacement would affect them.
Retain an owner, purpose and review cadence for each agent. Recent retrieval history is useful context, not a complete activity record: an agent may reuse a retained key or legitimately run infrequently.
Containment readiness is a launch prerequisite, not work to defer until the first incident.
A small release exercise
For an illustrative reporting agent:
- Connect a fictional or non-production credential to a trusted test runtime.
- Confirm retrieval of its approved key and denial of an unrelated credential.
- Verify the provider permissions needed for the report and rejection of an inappropriate operation.
- Confirm retrieval records contain the expected identity and outcome without the secret.
- Revoke retrieval and verify that a new request fails.
- Separately invalidate the provider credential and verify that retained copies no longer work.
Passing this exercise checks those paths; it is not a penetration test or proof that all production risks are resolved. Use the incident-response playbook for a suspected compromise and the credential lifecycle for ongoing reviews.
Gazebo can make retrieval access easier to manage across agents and services. Production readiness still depends on provider permissions, runtime controls, monitoring and the ability to act on an incident.