Researchers trained an open source AI search agent, Harness-1, that outperforms GPT-5.4 on recalling relevant information

Researchers from UIUC, UC Berkeley, and Chroma released Harness-1, a 20B open-source search agent built on gpt-oss-20B that scored 73% recall across eight complex search benchmarks, beating GPT-5.4 and leading open-source rivals. It externalizes search-session state into a structured environment to reduce “search amnesia” and was trained via Tinker. The model and weights are available under Apache