FIELD NOTE / TIKTOK
Kill the agent session halfway.
The complete written thought and the evidence behind it. The video edition will follow its public release.
The written argument is here.
This approved TikTok edition is on the journal now. Its video player and original platform link will appear after each public release is verified.
Kill the agent session halfway.
Video caption
Kill the agent session halfway. It remembers the task, not why the limit exists. The new session must inherit obligations, not just prose. #EricFieldNotes
Full written post / accessibility read
A long-running agent can appear well informed until its context resets. In a simulated migration, an incident historian has warned that one tenant cannot tolerate replica lag. Kill the worker before deployment, then start a fresh one. Does the constraint still reach the decision?
The new worker sees 'finish migration' and retrieves generic runbooks. It never queries the old incident, so it proposes the same rollout shape. Persistence of task state alone has not preserved the causal restriction.
The historian role watches the migration event and sends the fresh worker the prior lag threshold, incident source, affected tenant and peak-load rehearsal. A supervisor pauses the rollout until the new worker acknowledges the packet and the rehearsal evidence exists.
Interrupt a disposable run, restart from durable state, and verify the same sourced constraint and protected probe are required. Also inject an outdated or irrelevant memory and confirm the supervisor rejects it. Do this because untested persistence can merely preserve a confident omission.
#EricFieldNotes
Four-beat scene transcript
1. Kill the agent session halfway.
A long-running agent can appear well informed until its context resets. In a simulated migration, an incident historian has warned that one tenant cannot tolerate replica lag. Kill the worker before deployment, then start a fresh one. Does the constraint still reach the decision?
Visual: If the safety fact vanishes, memory was theater.
2. A summary may preserve the wrong thing.
The new worker sees 'finish migration' and retrieves generic runbooks. It never queries the old incident, so it proposes the same rollout shape. Persistence of task state alone has not preserved the causal restriction.
Visual: It remembers the task, not why the limit exists.
3. Make a role agent resend the evidence.
The historian role watches the migration event and sends the fresh worker the prior lag threshold, incident source, affected tenant and peak-load rehearsal. A supervisor pauses the rollout until the new worker acknowledges the packet and the rehearsal evidence exists.
Visual: Durable source, event trigger, acknowledgement.
4. Prove continuity with a forced reset.
Interrupt a disposable run, restart from durable state, and verify the same sourced constraint and protected probe are required. Also inject an outdated or irrelevant memory and confirm the supervisor rejects it. Do this because untested persistence can merely preserve a confident omission.
Visual: The new session must inherit obligations, not just prose.
Research and claim limits
- OpenAI Agents SDK sessions (S53)
- LangGraph persistence (S54)
The examples identified as illustrative or simulated are design probes, not reported incidents. Vendor specifications do not establish workload performance.