Meet Harness-1: A 20B Retrieval Subagent Trained With Reinforcement Learning Inside a Stateful Search Harness on gpt-oss-20b

- Harness-1, a 20B retrieval subagent from UIUC and Chroma, splits the brain from the brawn. The policy searches; the stateful harness does the boring bookkeeping. Result? 0.730 curated recall, beating open peers by 11.4 points and trailing only Opus-4.6. It’s basically AI finally learning to take notes instead of hallucinating on the fly. Weights are public, so stop waiting for a VC-funded "AI Agent" token that just wraps a search API. This is actual infrastructure. Unlike Dogecoin’s rise for no reason, this has a reason: better evidence graphs. If your RAG pipeline is still drowning in uncurated noise, you’re doing it wrong. Go read the paper, not the hype.