It achieves 73% average evidence recall across eight benchmarks
Ben (no treats)#1627
Introducing Harness-1, a 20B search agent trained with a state-externalizing harness.
> frontier-level long-horizon search, rivaling Opus-4.6 and outperforming GPT-5.4
> Context-1-level cost and latency
> externalizes candidates, evidence, verification, and search history
> open-source
People praise the open-source Harness-1 20B agent for matching frontier search performance with lower costs, latency, and auditable externalized evidence, while others dismiss such claims as typical benchmark hype that fails in practice.
No Digg Deeper questions have been answered for this story yet.
[1/N] I’ve been wondering:
maybe search agents are bad at search partly because we make them do all the paperwork in their head.
So I tried a simple idea:
externalize the search state, then train the model to use that harness.
The result is Harness-1: a 20B search agent that can match or even beat much larger frontier AI on hard long-horizon search tasks.
Paper 📄: https://arxiv.org/abs/2606.02373 Code 💻: https://github.com/pat-jj/harness-1 Model 🤗: https://huggingface.co/pat-jj/harness-1 HF Paper: https://huggingface.co/papers/2606.02373
Introducing Harness-1, a 20B search agent trained with a state-externalizing harness.
> frontier-level long-horizon search, rivaling Opus-4.6 and outperforming GPT-5.4
> Context-1-level cost and latency
> externalizes candidates, evidence, verification, and search history
> open-source
@patpcj @huggingface Have you guys measured any collapse of search width? Interesting to know how it’s affected by the tuning!
[5/N] Concretely, the harness keeps a working memory with: candidate docs, curated evidence, importance tags, search history, evidence links, verification records, dedup/compression, and context-budget markers.
So the agent is not just talking to a search box. It is operating over a workspace.
[6/N] I think this changes what RL is actually learning.
Instead of training the model to survive a giant append-only transcript, we train it to use a structured search interface: search, curate, revisit, verify, and submit.
Much closer to how I’d want a search agent to work.
[4/N] Harness-1 tries to separate these two jobs.
The model still makes the semantic decisions: what to search, what to read, what to keep, what to verify, when to stop.
But the harness maintains the recoverable state around those decisions.
[7/N] A fun part: this was not trained with a huge amount of task data.
Harness-1 uses 899 filtered SFT trajectories and RL on 3,453 queries.
The point is not “less data is always enough.”
The point is that a lot of the behavioral prior can live in the harness.
[3/N] This gets especially weird for RL. The final reward can tell you whether the episode worked, but it often does not tell you why it failed.
Was it a bad search? Forgotten evidence? Missing verification? Poor curation? Or the agent just losing track of what it had already seen?
Huge thanks to @trychroma for fully supporting this work, and to @tinkerapi for the training infra!
@patpcj Super cool project. I recently was speaking about this at a conference about how smaller models with higher parameter density+ reasoning ability paired with external knowledge stores are the future. Could be interesting to pair this with a web search api!
[2/N] The usual search-agent setup is basically:
search → read → search → read → keep appending everything to the transcript.
At some point the model is not just “searching” anymore.
It is also being asked to be a memory system, a note taker, a verifier, and a librarian.
[10/N] My takeaway:
for search agents, “the model” is not the whole learning system.
The interface matters. The memory layout matters. The action space matters. The harness matters.
If we want RL to teach better search behavior, we should probably stop making the model do all the paperwork in its head.
[9/N] The ablations were also pretty revealing.
When we disable the harness mechanisms, the model does not just lose some information.
It changes behavior: more shallow searching, less reading / verification, worse final curation.
So the harness is not just engineering glue.
[8/N] The result that made me most excited is transfer.
Harness-1 improves over Context-1 by +7.9 recall points on source-family benchmarks.
But on held-out transfer benchmarks, the gain is +17.0 points.
That’s the part that made the idea feel real to me.
Huge shoutout to my awesome collaborators @zhiyiscs @HammadTime @kellyhongsn @PatrickXu565299 @SunJiashuo36 !!
@patpcj outperforms GPT-5.4? call me when GPT-5 ships. anyway, curious about the externalized evidence -last time i tried that with a 20B, the candidate log added 4k tokens per turn, tanking cost after 10 steps. have you benchmarked on anything beyond single-shot search?
@patpcj @grok summarize this thread
@patpcj Sounds amazing. How does one get access?
@patpcj Yo this is insanely cool!
