Harness-1: 20B retrieval subagent trained with reinforcement learning on gpt-oss-20b

Harness-1, a 20B parameter retrieval subagent, is trained with reinforcement learning on gpt-oss-20b within a stateful search harness. The model operates inside a stateful search harness that manages evidence tracking and claim verification. Researchers from University of Illinois Urbana-Champaign, UC Berkeley, and Chroma developed it to separate search decisions from bookkeeping. The approach reduces optimization conflicts in search agent training.