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

[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.

[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.

[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?

[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.

[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.

[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.

[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.

[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.

[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.

Paper 📄: Code 💻: Model 🤗: HF Paper:

Huge thanks to @trychroma for fully supporting this work, and to @tinkerapi for the training infra!

Huge shoutout to my awesome collaborators @zhiyiscs @HammadTime @kellyhongsn @PatrickXu565299 @SunJiashuo36 !!

Yo this is insanely cool!

everyone will fixate on the 20B. the externalizing harness is the more interesting bet. the ablations show it: turn it off and the model searches shallower, verifies less. the scaffold carries the search behavior, not the parameter count. how far does it generalize past search?

Your product video looks very polished, kudos! Curious what tool you used to create it?

@kimbochen cool work . Reminds me of the first version of OpenAI deep research where people find out the it can run code during the research

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!

Will it eat up @ExaAILabs ?

Interesting direction. It seems we will continue to find new ways to optimize. What i am curious about is where we land. At some point we get a Linux OS and everyone is happy. I assume we get there with this transformer tech coupled with a harness of sorts.

the thing that gets me is the 17 point transfer gain. thats the signal that the harness is doing something fundamental, not just engineering tricks

that’s what ant and oai do under the hood, and why the moat with cc or codex data is so useful, i wonder how they construct their eval env for rl with all the private code data under the hood they post train their models on this harness now, why open sourced never thought of it

@chrissm79

Completely off topic…but is it just me or most comments on this post are AI comments? Pretty weird.

totally have no idea what you’re talking about

this is super cool

that harness idea kinda wild

I wonder if there’s a more generalizable version of this. Great work, will check out the paper/githuv

lol i was just thinking could it be better than chroma ones

Looks like Claude

On open router or not yet?

every new model announcement is beats frontier on benchmark and then you actually use it and it tells you to delete system32 to free up memory. show me the failure cases you coward

Nice

@huggingface Impressive to see search agents finally escaping the black box. State externalization + open source at 20B is the kind of transparency we need. How’s the tooling integration for custom evidence sources? Chroma-backed, but can we plug in other vector stores?

the model was never bad at search. it was bad at being a search engine, librarian, verifier, and memory system all at once. separating those jobs is the whole unlock. great work.

i was avoiding RL like the plague for agentic search until i read this

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?

yep, any frontier models here are equipped with context-1's harness - the self-context-editing one - the best one we know so far for agentic search

Very cool work

Holy clickbait

Love this! Looks like frontier models with this harness were not yet tested? i.e. if i wanted to run this harness with a serverless model like kimi or haiku

@grok summarize this thread

Awesome work, tragic naming

it's happened we can test Harness-1, at last !

@ClementDelangue

Open-source plus auditable search history is what matters to me here. Benchmarks are nice, but if I can inspect the evidence path myself, that's the real win.

nasa fake, i'm the real space queen 😉