This “state externalization” trick is pretty interesting—small models can also handle long-range retrieval. It’s worth following the open-source release of Harness-1.

CryptoWorld news: the 20B retrieval intelligent agent Harness-1 has been open-sourced. Researchers come from UIUC, UC Berkeley, and Chroma. The model uses an external state architecture, offloading the memory and organization work during retrieval to the environment side, enabling non-frontier-scale models to achieve near state-of-the-art performance in long-range search tasks with very little training data. On 8 retrieval benchmarks covering web pages, finance, patents, and multi-hop question answering, Harness-1 achieves an average filtered recall rate of 0.730—11.4 percentage points higher than the second-best open-source retrieval sub-agent—showing that explicitly modeling retrieval state helps the model learn more transferable search strategies.