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 unveiled Harness-1, a 20B-parameter open-source search agent based on OpenAI’s gpt-oss-20B. It scored 73% average recall across eight complex search benchmarks, beating GPT-5.4 and leading open-source rivals, with code and weights released under Apache 2.0.

Harness-1 suggests open-source agents are quickly closing the gap with proprietary systems on demanding enterprise retrieval tasks like financial filings, patents, and multi-hop research. Its use of Thinking Machines’ Tinker also points to a broader trend: better training infrastructure is making specialized autonomous agents easier to build and share.