Coding agents need typed context surfaces, not one flat context window
Puzzleheaded_Box2842 · reddit · 2026-10-09
The author argues that flattening source files, chat history, configs and retrieved text into a single context makes coding agents unreliable for tasks needing exact values, structured relationships, reusable procedures, or durable facts — each demands a different access pattern (hybrid retrieval for documents, SQL for structured data, graph traversal for relationships, versioned instructions for procedures).
The proposed architecture exposes typed context surfaces instead of one vector index:
- KB for documents/notes with hybrid retrieval
- DB for structured records and SQL
- Graph for entities and bounded traversal
- Skills for reusable versioned procedures
- Memory for scoped, durable facts
Requests route to the relevant surface, evidence keeps provenance, and reads/writes are separated — updates return explicit receipts with source, revision, and results. Implemented in the open-source Codex MCP plugin OpenDCAI/DataMind.
More from coding & agent
- PiDurableKit: Open-Source Durable Agents for Apple Platforms That Survive iOS Kills — aigclink · 2026-10-09
- PiChat: An Autonomous Agent Living in Your iPhone with Durable, Resumable Execution — aigclink · 2026-10-09
- One prompt and a track: Opus 5.5 autonomously ships a finished product in 12 hours — FinanceYF5 · 2026-10-09
- The expensive part of AI video isn't generating it, it's regenerating it — Khalizo · 2026-10-09
- Production AI agents cut loan underwriting from 6 days to under 20 minutes — alex_verem · 2026-10-09
- Same 15-second 3D brief, two agents, four tools: an eight-run comparison — tobowers · 2026-10-09