Deliver context when decisions need it.

Day 19 · 2026-10-12 · LinkedIn

Short video caption

RAG helps when a worker knows what to ask. A security-context agent with durable, reviewed memory can watch a revocation change and inject a small sourced tenant constraint before the decision. Require recipient acknowledgement and a relevant test. Illustrative system; evaluate missed constraints and interruption cost against retrieval-only baselines. #EricFieldNotes

Full written post / accessible read

Strong agent teams should not flood every worker with the organization's entire memory. They should deliver the right constraint before a decision can violate it. Retrieval helps with questions the worker knows to ask. It can miss the fact the worker never thought to query.

Imagine a worker changing session handling. It retrieves general auth docs, but never asks about a tenant whose contract requires immediate revocation. The exception exists in the organization's history; the worker does not know to search for it.

A security-context agent keeps durable, reviewed memory of auth exceptions. When the code change touches revocation, it sends the worker one bounded message: the constraint, its source, its owner and the test it requires. The supervisor records whether the worker acknowledged it.

Subscribe specialist agents to consequential events. Inject only relevant facts with a source, freshness limit and required action; require acknowledgement before the protected decision, then test the outcome. Do this because passive retrieval cannot fetch a constraint the worker never names.

#EricFieldNotes

Evidence and boundary

On-screen boundary: ILLUSTRATIVE AGENT SYSTEM. The sources below support documented mechanisms and specifications; illustrative scenarios are not presented as measured incidents.

• OpenAI Agents SDK/Codex orchestration example (S07)
• OpenAI Agents SDK sessions (S53)
• LangGraph persistence (S54)