Open-Source AI Search Agent Harness-1 Outperforms GPT-5.4 on Recall Benchmarks Despite Training on Fraction of Rival Data

Researchers from UIUC, UC Berkeley, and Chroma unveil Harness-1, a 20-billion parameter open-source AI search agent that outperforms GPT-5.4 on recall benchmarks while training on a fraction of rival data, achieving 73% recall by offloading memory management to a structured external environment.

• Researchers from UIUC, UC Berkeley, and Chroma unveil Harness-1, a 20-billion parameter open-source AI search agent that scores 73% on recall benchmarks, outperforming GPT-5.4 at 70.9% and beating the next best open-source competitor by 11.4 percentage points.
• Harness-1 achieves its breakthrough performance by externalizing search state into a structured 'harness' environment that manages working memory, evidence curation, and verification records, freeing the model from the cognitive overload of tracking everything in its own context window.
• Trained on just 899 supervised fine-tuning trajectories and 3,453 reinforcement learning queries, Harness-1 demonstrates radical data efficiency compared to rivals requiring hundreds of thousands of training items, and is released under the permissive Apache 2.0 license for immediate commercial use.