Quick answer: Multiple agents introduce delegation and credential-propagation paths. A named retrieval profile for each worker can make its credential requests distinguishable, but does not automatically isolate its runtime, narrow a shared provider key or create a complete trace of a delegated task.
For an overview of identities and enforcement layers, see IAM for AI agents. This guide concentrates on what an orchestrator and its workers can actually reach.
Map the credential flow first
An orchestrator may start workers as processes, call them over a network or let them use shared tools. These designs have different trust boundaries. A subprocess may inherit environment variables unless the launcher deliberately restricts them; a remote worker receives only what the caller and its hosting system provide. Neither arrangement automatically proves that one agent cannot access another's token or a privileged shared tool.
An illustrative risky flow is:
Orchestrator holds a broad provider key
-> starts a worker with that key in its environment
-> sends another worker the key in a task payload
-> both workers can use the provider key's actual permissions
More copies create more places to expose or misuse the same key; they do not multiply its provider permission set by the number of agents. A single worker holding that key already has the key's effective authority.
Give each worker only the authority it needs
Prefer passing a task description without a provider key. Give a worker an appropriate identity and limited tools in an execution context that protects those boundaries. Then decide whether it should use native provider workload identity, retrieve a configured key or call an application tool that performs a narrow operation on its behalf.
When using Gazebo, separately authenticated workers may have separately configured retrieval policies and history. An approved retrieval returns the underlying provider key to the worker runtime. If two workers retrieve the same key, their downstream provider authority remains the same. For different provider capabilities, use appropriately restricted provider credentials or an execution layer that actually enforces the difference.
Protect the bootstrap tokens too. If the orchestrator can read every worker's token, filesystem or tool, a naming convention cannot stop it from indirectly using a privileged worker. Distinct processes and distinct profiles are useful only with appropriate access to their credentials and tools.
Delegation is not self-authorization
A worker asking a more privileged worker to perform an operation should not make that operation authorized. The executing tool needs to know the authenticated caller, the delegated task and the applicable policy. Bind a human approval to the proposed action before execution when a second decision is required, and reject requests that exceed the authority actually delegated.
This is a design for the executing system; Gazebo's OAuth approval of named keys and its credential-release checks do not themselves implement per-action approval or a delegation protocol. A reviewer that looks at results after execution can detect problems, but cannot retroactively prevent a charge or deployment.
Example: planner, executor, reviewer
Consider three workers in a single workflow:
- Planner: receives task context and creates a plan. It has no provider credential and makes no credential-retrieval request. Its planning activity therefore does not appear as a Gazebo credential-access event.
- Executor: performs a permitted provider operation using a provider identity or a suitably restricted key. Its Gazebo retrieval identity, if used, determines whether it may receive that configured key—not which later Stripe payment or GitHub PR operation the provider will allow.
- Reviewer: examines a proposal before a high-impact operation if prevention is intended, or examines its result afterward for detection. Read-only access must be enforced by the actual provider key and tools, not by a
reviewerprofile name alone.
This is an example architecture, not a description of an automatically provided Gazebo pipeline. To correlate all three stages, the application needs to record its own task/run identifiers and links between worker invocations. Protect that context against untrusted callers inventing an authorization claim.
Test the boundaries, not just the diagram
Check that each worker can do its authorized task and is denied a selected out-of-scope retrieval. Separately test the provider operation with the actual key: a Gazebo request declaring GET does not make a broad provider key read-only. Check whether workers can read one another's environment or use a shared tool that already holds more authority.
When a worker loses access, revoke its retrieval identity to stop new release through that route. Isolate its runtime and assess keys it already received. If the same provider key is used elsewhere, invalidating it at the provider may affect those consumers; leaving it active may let a compromised worker continue using its copy. Neither outcome is guaranteed away by separate profile names.
Correlate evidence without overclaiming
Gazebo can record which authenticated retrieval identity requested which key and the retrieval outcome. Provider logs may identify the provider key, not the worker that retrieved it. Your runtime can add task/run correlation and execution outcomes where appropriate; minimise sensitive payloads and protect the resulting records.
A failed provider operation is not necessarily a denied Gazebo retrieval, and a worker that never retrieved a key will not have a credential-access event just because it took part in the task. Investigate both evidence sources before adjusting a permission.
For a task-to-policy worksheet, see AI agent permissions. For offboarding and incidents, see AI agent credential management.