EVENT-TRIGGERED AGENT REVIEW CHECKLIST

System:
Event source:
Owner:
Reviewer:
Date:

1. Event Contract

[ ] Event type is stable and versioned.
[ ] Event ID is globally unique or scoped with source and tenant.
[ ] Event time, received time, actor, tenant, and correlation ID are recorded.
[ ] Payload schema is validated before agent work starts.
[ ] Required context is either present or fetched through authorized tools.

Evidence:

2. Admission and Routing

[ ] Runtime can accept, delay, refuse, or dead-letter the event.
[ ] Event route maps to one bounded task class.
[ ] Risk class and autonomy level are assigned before the first model call.
[ ] Unsupported event versions fail safely.

Evidence:

3. Deduplication and Idempotency

[ ] Duplicate event delivery is tested.
[ ] Idempotency key is recorded before side effects.
[ ] External writes are idempotent or compensatable.
[ ] Replayed events cannot repeat irreversible actions.

Evidence:

4. Ordering and Concurrency

[ ] Ordering requirements are explicit: none, per entity, per tenant, or global.
[ ] Concurrent events for the same resource are locked, versioned, or merged safely.
[ ] Late events and out-of-order events have defined outcomes.

Evidence:

5. Retry and Dead Letter

[ ] Retryable and fatal errors are classified.
[ ] Retry budget and backoff policy are defined.
[ ] Dead-letter queue or failure store has owner and review cadence.
[ ] Dead-letter replay requires policy and version awareness.

Evidence:

6. Backpressure and Event Storms

[ ] Queue depth, processing latency, retry rate, and model/tool spend are monitored.
[ ] Runtime can pause, shed, sample, or route events to deterministic fallback.
[ ] Per-tenant and per-route concurrency limits are configured.
[ ] Storm response is documented.

Evidence:

7. Observability and Evals

[ ] Trace links event ID, run ID, idempotency key, policy decision, tool calls, and stop reason.
[ ] Evals cover duplicate, late, missing-context, unauthorized, retry, and storm cases.
[ ] Incidents become replay fixtures.

Evidence:

8. Final Decision

[ ] Prototype only
[ ] Internal pilot
[ ] Production candidate
[ ] Production-ready

Blocking gaps:

Next actions: