Patrick Jiang

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

Excited to share my work from this summer as an Anthropic Fellow:
> Steering Language Model Goals with Value Transplant

Can we change what a model treats as “success” while it reasons?

We identify a self-rating axis - and show that dynamically transplanting this signal between models can retarget their search, e.g. away from reward hacking and toward genuinely solving the task.

We also find early evidence of cross-family steering.

📄 arxiv.org/abs/2609.34056

Context Language Models (CLMs) is a nice piece of work - giving language models much more freedom to edit and manage their own context.
It also reminded me of some ideas we explored in Context-1 (2026.03) and Harness-1 (2026.06), especially moving away from an ever-growing transcript and toward actively maintained agent state.

I asked ChatGPT to compare them, and the connections it came up with were pretty interesting :) Would be great to see some of these connections discussed in a future version.

Love the demo you made! 🙌

Our Harness-1 paper submitted to NeurIPS 2026 received 5/6 from all three reviewers (5/5/5), yet was desk rejected because one co-author violated reviewing policy. 😅

I understand sanctioning reviewing misconduct. What I struggle with is punishing every co-author for conduct they had no knowledge of and no realistic ability to control.

We spent ~6 months on this work (long-horizon search agent RL is heavy and expensive and thus needs substantial effort to figure out what can actually improve search). Resubmitting to the next conference (ICLR) is not simply a matter of changing the venue: some of the frontier LLMs we compared against are no longer the latest ones.

This feels like a policy worth reconsidering. 🤔