Harness-1: Reinforcement Learning for Search Agents with State-Externalizing Harnesses
Pengcheng Jiang, Zhiyi Shi, Kelly Hong, Xueqiang Xu, Jiashuo Sun, Jimeng Sun, Hammad Bashir, Jiawei Han
arXiv:2606.02373 · 59▲ · cs.AI, cs.CL, cs.IR
View on arXiv →Premium readers get an interactive explainer for this paper — a figure you can poke at, not just read.
What it is
Harness-1 is a 20B search agent (built on gpt-oss-20b) trained with reinforcement learning inside a stateful harness that moves search bookkeeping out of the model and into the environment. The environment maintains a candidate pool, an importance-tagged curated set, an evidence graph, verification records, and deduplicated/compressed observations, while the policy only decides what to search, which documents to keep, what to verify, and when to stop.
Why it matters
It reframes retrieval-agent design around the harness (the interface layer between the model and its tools) rather than piling all state tracking into the prompt transcript. For practitioners building RAG or search pipelines, the concrete lesson is that externalizing routine state into the environment lets a small model match or beat much larger frontier models on evidence collection, and the same fixed model (GPT-5.4) gained about 4 recall points just by swapping to a richer harness with no retraining.
Practical takeaway
If you build multi-turn search agents, consider maintaining an explicit editable state (candidate pool, curated set with confidence tags, verification cache) in your harness instead of relying on the model to reconstruct it from a growing transcript. The code is promised open-source, so you can watch for the released harness and RL recipe to adapt the pattern.
Key result
0.730 average curated recall across eight retrieval benchmarks (web, finance, patents, multi-hop QA), which is +11.4 points over the next best open agent (Tongyi DeepResearch 30B); it beats GPT-5.4, Sonnet-4.6, and Kimi-K2.5 under the same protocol, though Opus-4.6 still leads on average. Caveat: results are point estimates with no confidence intervals, and frontier models were run under a different (Context-1) harness rather than Harness-1's own.
Subscribe
Get the next issue.
Free. One email a week. Unsubscribe any time: no account, no dark patterns.