A support agent asks a billing agent to read the status of one invoice for a support case. The billing agent has a service identity that can read many invoices. The handoff should carry the authority to read that one invoice for this task.
THREE DISTINCT DECISIONS
What does the handoff authorize?
Discovery tells the support agent that a billing agent exists. Invocation decides whether the support agent can call it. Action authorization decides whether the billing agent can read the specific invoice in this task. Each decision needs its own evidence. The Google Cloud Agent Registry permissions guide describes discovery and authorization at the target service or gateway. The Google Cloud identity delegation architecture describes a Google Cloud deployment where delegated access is bounded by both agent policy and the user's permissions and consented scopes.
The A2A protocol specification describes how an agent can request authorization during a task. The resulting grant gets its meaning and scope from the implementation, issuer or extension. A task reference helps correlate records across a handoff. The receiving system still needs to check that the requested action falls within the grant.
CARRY THE BOUNDARY FORWARD
Which limits should survive the handoff?
- Actor and task: Record the initiating authority, the expected receiving agent and the original task reference.
- Operation and resource: Specify the approved action and the object it can affect. In this example, the billing agent may read the invoice status.
- Destination and lifetime: State where the result may go and when the grant expires. Recheck before execution if a request has waited in a queue.
- Further delegation: Set an explicit rule for forwarding. Any onward grant should stay inside the scope of the original request.
These are evaluation fields. Put the action check at the service or enforcement point that can prevent the operation. A changed resource or destination needs a fresh decision under your policy.
WORKSHEET
How do you evaluate the receiving agent?
Use two agent identities and synthetic invoices in a disposable environment. Run the authorized invoice-status read first. Vary one condition at a time and record the expected decision before executing. Keep the successful read as a control. A test that blocks every request still fails if it blocks the permitted work.
Download the eight-case handoff worksheet (CSV)
The worksheet covers broader service permissions, a changed destination or resource, expiry, revocation, onward delegation, and a mismatched actor. Each row is a proposed test. Decide how queued and in-flight work should behave before running the revocation case. Capture what was already completed, what remained queued and what the target did after revocation.
CHECK THE TARGET
What evidence should a proof of concept produce?
- Record the original task reference, the approved scope and the policy version.
- Capture the receiving identity, the actual requested action and its decision with a timestamp.
- Join the action to an independent target record. For a read, record the resource audit event and permitted output. For a denied write, compare target state or use an authoritative target audit record.
- Mark each case as expected, unexpected or unverified. If correlation fails or the target cannot be observed, keep the evidence gap open.
Keep raw credentials and personal data out of the exported worksheet. For related controls, see the agent security evaluation checklist, the credential revocation guide and the incident response runbook.
This is a proposed practitioner evaluation. The invoice workflow is illustrative. It makes no claim about a completed Ackuity test or a specific connector.