Harness-1: Reinforcement Learning for Search Agents with State-Externalizing Harnesses

Researchers introduced Harness-1, a 20B open search agent that uses state-externalizing harnesses to manage documents, evidence, verification, and context outside the model.

It achieved strong average curated recall across eight retrieval benchmarks, outperforming the next best open search subagent and showing especially large gains on transfer benchmarks.

Ablation studies found that removing the harness reduced recall and changed model behavior, underscoring the value of structured state management.

The results suggest search-agent performance can improve when memory and bookkeeping are handled externally, easing model burden and improving transfer to new retrieval tasks.