harness-1

Why we included this project

Harness-1 bundles a 20B search agent's trained weights with the exact training recipe that produced it, so you can either run the model or reproduce the pipeline. The checkpoint serves locally through vLLM, and the stateful retrieval harness keeps candidate documents, evidence links, and budget-aware context explicit across turns, letting the policy decide what to search and when the evidence is enough. The SFT trajectories, RL training records, and retrieval corpora are published in the repo, which matters if you want to adapt the approach rather than take a closed model on faith. Teams building agentic tools where multi-step retrieval and evidence verification need to be inspectable will find the design unusually transparent. One honest caveat: full evaluation expects a compatible retrieval backend, so it is not a plug-and-play self-contained demo.