Build
Production operation
Roll out, observe and recover multi-step workloads
Agent reliability depends on both orchestration and model access. Test their failure boundaries together before customer traffic.
Shadow policy first
Run representative workflows with Router policy in shadow, review would-deny evidence, then promote one internal tenant to enforce with a tested rollback.
Golden fixtures
Cover text, streaming, tools, structured outputs and multimodal paths for every approved provider used by the workflow.
Recovery
Checkpoint before side effects, distinguish incomplete model output from tool failure and require explicit operator or product logic before replay.
Incident operation
Expose build identity, request correlation, provider health and workflow state. Practise disabling a route, rotating a key and rolling back policy without importing product code into Router.