Harness-1: 20B retrieval agent trained with RL inside stateful search
UIUC and Chroma released Harness-1, a 20B retrieval subagent trained via reinforcement learning within a stateful search framework that tracks candidates, curated sets, evidence graphs, and verification records.
Achieves 0.730 average curated recall across eight benchmarks — 11.4 points above the next open subagent, trailing only Claude Opus 4.6.
Weights and harness code are open-sourced.