Do not pour the whole memory into every agent.

Video caption

Do not pour the whole memory into every agent.

A fluent summary without a source can be dangerously stale.

Deliver one claim before dispatch and recheck the record.

Inject timely evidence; let the target decide permission.

Narration uses Eric's authorized AI voice clone.

#EricFieldNotes

Full written post / accessibility read

A worker is preparing a tenant change. Somewhere in years of notes is a current customer exception. Dumping the archive into the prompt makes the decisive fact harder to notice; relying on one search query can miss it. A steward can watch for this exact action and deliver a narrow packet.

The steward says this customer may be moved. When was that true? Under which policy version? Was it later revoked? In the visual, the source record and expiry travel with the short message. Without them, the worker has a persuasive sentence and no safe way to revalidate it.

The supervisor subscribes to a proposed production effect, finds the relevant policy owner record, and sends a compact claim plus a question if evidence conflicts. The target service reads current eligibility anyway. Test omitted, stale and contradictory packets in a disposable fixture.

Keep the steward's memory and reasoning separate from the worker's short context. Send only the evidence required for the imminent decision, with source and expiry. Require a current service check because better context reduces missed facts but cannot itself prove the action is authorized.

Narration uses Eric's authorized AI voice clone.

#EricFieldNotes

Four-beat scene transcript

1. Do not pour the whole memory into every agent.

A worker is preparing a tenant change. Somewhere in years of notes is a current customer exception. Dumping the archive into the prompt makes the decisive fact harder to notice; relying on one search query can miss it. A steward can watch for this exact action and deliver a narrow packet.

Visual: The relevant exception matters at one decision point.

2. The packet needs provenance.

The steward says this customer may be moved. When was that true? Under which policy version? Was it later revoked? In the visual, the source record and expiry travel with the short message. Without them, the worker has a persuasive sentence and no safe way to revalidate it.

Visual: A fluent summary without a source can be dangerously stale.

3. Trigger at consequence.

The supervisor subscribes to a proposed production effect, finds the relevant policy owner record, and sends a compact claim plus a question if evidence conflicts. The target service reads current eligibility anyway. Test omitted, stale and contradictory packets in a disposable fixture.

Visual: Deliver one claim before dispatch and recheck the record.

4. Use active context, bounded authority.

Keep the steward's memory and reasoning separate from the worker's short context. Send only the evidence required for the imminent decision, with source and expiry. Require a current service check because better context reduces missed facts but cannot itself prove the action is authorized.

Visual: Inject timely evidence; let the target decide permission.

Research and claim limits

• OpenAI Agents SDK handoffs (S207)
• Anthropic contextual retrieval technical note (S209)
• OWASP Authorization Cheat Sheet (S229)

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