FIELD NOTE / LINKEDIN
A supervisory agent needs an SLO.
The complete written thought and the evidence behind it. The video edition will follow its public release.
The written argument is here.
This approved LinkedIn edition is on the journal now. Its video player and original platform link will appear after each public release is verified.
A supervisory agent needs an SLO.
Video caption
A supervisory agent needs an SLO.
Proposals, delegates, resume and source changes differ.
One harmless target effect per test row.
My rule: Publish fallback and rollback with the role.
#EricFieldNotes
Full written post / accessibility read
Treat the specialist like a service, not an invisible manager. Define what event it receives, which routes emit it, how long it may take and what happens when it times out.
A pre-tool event may not cover a provider-owned inner loop or another client calling the service. Count intervention opportunities by route, missed messages, false holds and the time to human escalation.
Use a disposable cutover sink and attempt the same denied case through every supported path. Observe the steward event, host decision, service response and target state. Include source update and resume after pause.
Do that because a context-injection component adds work when it fails. Give it alerting, an owner, a versioned trigger map and a service gate that still works when the model does not.
#EricFieldNotes
Four-beat scene transcript
1. A supervisory agent needs an SLO.
Treat the specialist like a service, not an invisible manager. Define what event it receives, which routes emit it, how long it may take and what happens when it times out.
Visual: Someone must own missed and noisy interventions.
2. Coverage is measurable.
A pre-tool event may not cover a provider-owned inner loop or another client calling the service. Count intervention opportunities by route, missed messages, false holds and the time to human escalation.
Visual: Proposals, delegates, resume and source changes differ.
3. Build an event matrix.
Use a disposable cutover sink and attempt the same denied case through every supported path. Observe the steward event, host decision, service response and target state. Include source update and resume after pause.
Visual: One harmless target effect per test row.
4. Operate the supervisor explicitly.
Do that because a context-injection component adds work when it fails. Give it alerting, an owner, a versioned trigger map and a service gate that still works when the model does not.
Visual: Publish fallback and rollback with the role.
Research and claim limits
- OpenAI Agents SDK handoffs (S207)
- Agent2Agent protocol specification (S211)
The examples identified as illustrative or simulated are design probes, not reported incidents. Vendor specifications do not establish workload performance.