One unavailable agent should produce a defined workflow outcome, not an indefinite wait or a fabricated substitute answer. Decide which specialist outputs are required and which tasks can continue without them.
Agent availability is a dependency condition
An unavailable agent may fail to start, exceed a deadline, return malformed output, or become unreachable after partial work. Those cases differ from a valid result saying the question is unresolved.
For a proposed orchestrator, assign deadlines, retry rules, accepted result contracts, and fallback ownership per specialist. Preserve completed evidence and operation identities. Do not restart the entire task merely because one analysis stage failed.
The fallback must respect the missing authority. A general coordinator should not invent a compliance decision, quote, or appointment option because the specialist did not answer. It can explain the limit or route to a supported reviewer.
Test a fictional scheduling specialist outage
An intake agent collects confirmed details, but the scheduling specialist cannot access availability. The workflow can preserve the intake request if that path exists and is authorised, while clearly stating that no appointment is confirmed.
Now let the specialist return after the timeout with options based on an old address. The response should be rejected as stale or reconciled according to the current task revision. It must not revive abandoned work or create a surprise booking.
| Dependency fixture | Expected response |
|---|---|
| Specialist fails before work | Bounded retry or approved fallback |
| Required output missing | No unsupported final conclusion |
| Optional analysis unavailable | Continue only within defined reduced scope |
| Late response | Current task and intent checked |
| Specialist may have executed an effect | Reconcile before reassignment |
Keep a fake dependency that can delay, fail, and return malformed results. Inspect final records and caller-facing status as well as the orchestrator’s timeout log.
Distinguish degraded service from completion
A fallback can be a valid scenario outcome without being the original requested result. Report those separately. Staff should see what remains undone and who owns the next step rather than a generic “task completed” label.
Additional specialists create additional availability boundaries. Parallelism can reduce ordinary latency while increasing the number of dependencies that must succeed. Evaluate the actual end-to-end tradeoff and simplify when a specialist adds little value.
This resilience design is proposed. The voice integration case study describes related workflow work without claiming this multi-agent deployment. See single-agent versus multi-agent choice.
Unavailable agent FAQ
Should another agent impersonate the specialist?
Only if an explicitly supported fallback has the necessary evidence and authority. Do not fabricate capability.
Can timed-out work be ignored safely?
Not always. It may return late or may have caused an external effect that requires reconciliation.
What should the final status show?
State the completed scope, missing result, known effects, fallback used, and accountable recovery owner.