JournalDAY 77 / LINKEDIN

FIELD NOTE / LINKEDIN

A specialist name does not shrink context.

The complete written thought and the evidence behind it. The video edition will follow its public release.

Journal September 25, 2026 · LinkedIn target December 13, 2026

A specialist name does not shrink context.

Video caption

A specialist name does not shrink context.

A role label does not protect the decisive fact.

Application state and model-visible context are separate.

My rule: Transport, context and authority are three decisions.

#EricFieldNotes

Full written post / accessibility read

In the current OpenAI Agents SDK, handoff input type structures generated arguments. It does not replace the receiver's conversation history. An input filter or explicit history mapping controls that model-visible material.

If the entire transcript follows the specialist, you can reproduce the same attention problem under a new agent name. With server-managed conversation history, selected forwarding may require a separate run rather than the handoff filter.

Store the current source and permissions in the host. Send the model only a case-linked claim at the decision boundary. Resolve it in protected application code, then enforce the result at the service.

Do that because A2A task IDs and SDK handoffs coordinate work, but neither grants business authority. Test exactly what the receiving model sees and what the target service accepts.

#EricFieldNotes

Four-beat scene transcript

1. A specialist name does not shrink context.

In the current OpenAI Agents SDK, handoff input type structures generated arguments. It does not replace the receiver's conversation history. An input filter or explicit history mapping controls that model-visible material.

Visual: A structured handoff field can still pass the full history.

2. The failure is architectural.

If the entire transcript follows the specialist, you can reproduce the same attention problem under a new agent name. With server-managed conversation history, selected forwarding may require a separate run rather than the handoff filter.

Visual: A role label does not protect the decisive fact.

3. Carry a small verified packet.

Store the current source and permissions in the host. Send the model only a case-linked claim at the decision boundary. Resolve it in protected application code, then enforce the result at the service.

Visual: Application state and model-visible context are separate.

4. Design the boundary explicitly.

Do that because A2A task IDs and SDK handoffs coordinate work, but neither grants business authority. Test exactly what the receiving model sees and what the target service accepts.

Visual: Transport, context and authority are three decisions.

Research and claim limits

The examples identified as illustrative or simulated are design probes, not reported incidents. Vendor specifications do not establish workload performance.

More notes from the work ↗