What's this about?
LinearUncle posted a study list of production open-source agent "harnesses" — the scaffolding that turns a raw model into a working coding agent — and suggested letting AI mine them for features you don't know how to build. Baoyu's advice: go deep on one (pi-mono) rather than skimming all, and he notes the recently-leaked Claude Code source is the heavyweight the list is missing.
What exactly is a "harness"?
The code around the model that makes it an agent: the loop that feeds it tools, manages context, parses its tool calls, retries, and enforces acceptance criteria. The model is the engine; the harness is the car. Most of the quality difference between agents lives here, not in the base model.
What's the community saying?
355 + 468 likes.
- "Study one deeply beats skimming many" is the resonant note — harness design rewards depth.
- The leaked Claude Code source keeps coming up as the reference implementation people actually want to read.
- Fits the week's theme of harness-building as a distinct craft (see also Dwarkesh's compute essay and everyone's multi-model workflows).