Member-only story
Knowledge Layers for Agent Harnesses: Context Your Agents Can Discover on Their Own
When a team runs AI agents against its repositories at any real volume, the context an agent starts with is whatever its harness loads at the beginning of a session. Everything else, the reason a service retries three times, the postmortem that explains why the queue backed up last winter, the convention that this team opens an MR for config changes but pushes everything else to main, lives in people’s heads, in a wiki, in an incident report, or in a chat thread somewhere. A human pilot fills those gaps as they go. A headless run in CI has nobody to ask.
It’s not rare to find a team re-explaining the same background to an agent session after session, or watching a scheduled run go down a path that any engineer on the team would have steered it away from. The agent didn’t lack intelligence, it lacked reach.
This article is about closing that gap with a knowledge layer: curated context that agents can find on their own, piloted or not. It’s a survey of what people are doing as of October 2026, the options and their trade-offs, and an experiment I’ve been running on my own. These are as much notes to myself as anything else, so I’ll stick to documenting what exists and what I’ve tried rather than prescribing.
