Classification: reliability method based on two scoped published observations; no production throughput or success rate is claimed.
View the full infographic → · View this article
An agent job is not complete because a process is alive, a provider streamed output or the model said “done.” Define acceptance before execution and verify a durable artifact against that definition.
Write an acceptance contract
State the expected artifact, required fields, source boundary and deterministic checks. If judgement is needed, use an explicit rubric and identify when a person must review. Keep the check independent from a heartbeat or status message.
Track the stages separately
- Process is scheduled and alive.
- Provider call is pending, streaming, complete or failed.
- Workflow stage has emitted a checkpoint or artifact.
- Verifier evaluated the artifact.
- Job was accepted, partially accepted or rejected.
If a callback was sent but no acknowledgement arrived, preserve its identifier and reconcile delivery before resending. If a model result is missing, do not infer success or failure solely from a heartbeat. Record the uncertainty and inspect the attempt.
Connect claims to evidence
Post #42 documents one successful bounded canary with provider usage; it does not establish general success rates. Post #53 documents a reported stall whose exact earlier outcome is unknown; a later bounded check was separate. They show why liveness, provider activity and acceptance should not collapse into one status.
Use the agent rules builder for progress and verification safeguards, and read the measurement method before comparing runs.