# Harness Learning: AI Agents That Rewrite Their Own Harness

> A paper trains a small model to rewrite an agent's harness from feedback; Musk compared it to Tesla FSD. What it shows and what it does not.

- Author: Nitish Garg, Founder & CEO, CellCog
- Published: 2026-10-12
- Canonical (HTML): https://cellcog.ai/blog/harness-learning-test-time-adaptation/
- Section: Guides / Choosing a platform
- Publisher: CellCog (https://cellcog.ai), the AI employee platform. Blog index for agents: https://cellcog.ai/blog/llms.txt

## Key points

- A September 28 arXiv paper trains a proposer model to rewrite an agent's harness, the program around the model, using execution feedback from the task at hand.
- No weights change at test time. The paper treats harness revisions as the analogue of weight updates.
- On 21 held-out Reasoning Gym families, the trained 4B proposer raised mean single-step scores from 0.32 to 0.62 and beat its 35B teacher on average.
- Trained on HotpotQA, the proposer transferred to MuSiQue and 2WikiMultihopQA; on MuSiQue, ten rounds of revision reached 0.27 exact match versus 0.15 for independent revisions.
- Pedro Domingos shared it as the model swallowing the harness; Elon Musk compared it to Tesla FSD. In the paper the harness stays an explicit program a model writes.
- It is a preprint on 4B and 8B models doing puzzles and question answering, with MASS and Evo² posted in the same two weeks; none is peer-reviewed yet.

## At a glance

- **What is harness learning?** A method that trains a proposer model to revise an AI agent's harness, the program organizing its model and tool calls, using execution feedback, with no weight updates at test time.
- **What did it achieve?** On 21 held-out Reasoning Gym families, mean single-step scores rose from 0.32 to 0.62, and the 4B proposer beat its 35B teacher on average.
- **Did the model replace the harness?** No. In the paper the harness remains an explicit program; a model rewrites it while both models' weights stay fixed.

**A paper from researchers including Ruslan Salakhutdinov and Daniel Khashabi trains a small model to rewrite an AI agent's harness, the program that decides how the agent calls its model and tools, using feedback from the task in front of it.** It changes no model weights at test time. On 21 held-out Reasoning Gym task families, its trained 4B proposer raised the mean single-step revision score from 0.32 to 0.62. The paper, [Harness Learning Enables Generalizable Test-Time Adaptation](https://arxiv.org/abs/2609.35738), was posted to arXiv on September 28, 2026. It went viral on October 11 when Pedro Domingos shared it with one line: "the model swallows the harness".

Elon Musk replied three hours later: "Same thing happened with Tesla FSD. The model ate more and more of the C++ code, until there was nothing left." When we read X at 06:47 UTC on October 12, the Domingos post had 325,559 views and Musk's reply 231,833. Here is what the paper did, what its numbers say, and why "the model swallows the harness" is a reading of the paper rather than its claim.

## What the paper does

The authors start from a definition: "A language-model agent is jointly defined by its model and its harness". The harness is the executable program around the model: which calls it makes, which tools it uses, how information flows between steps. Different tasks want different harnesses, so the paper trains a second model, the proposer, to revise the solver's harness after watching it run.

They treat this as meta-learning over programs, "with harness revisions playing the role of weight updates in gradient-based adaptation". The proposer is first fine-tuned on revisions written by a 35B teacher, then trained with reinforcement learning, rewarded by how well each revised harness performs. At test time it adapts the harness to a new task "without performing any parameter-space update".

## The numbers

*Table: Harness learning, as reported in the paper (arXiv 2609.35738v1, September 28, 2026)*

| Setting | Models | What it found |
|---|---|---|
| Reasoning Gym, single revision | Qwen3.5-4B proposer and solver | On 21 held-out task families, mean score rose from 0.32 with the base proposer to 0.62 after supervised fine-tuning plus RL |
| Reasoning Gym, against the teacher | 4B proposer vs its 35B teacher | The trained 4B proposer beat the 35B teacher on average |
| Multi-hop QA, transfer | Qwen3-4B proposer, Qwen3-8B solver | Trained on HotpotQA, it transferred to MuSiQue and 2WikiMultihopQA, which it never trained on |
| MuSiQue, repeated revision | Same | 0.15 exact match averaged over independent revisions, versus 0.27 after ten rounds of revision |

The figure that matters most is the transfer. The proposer learned to revise harnesses on some tasks and kept improving them on tasks it had never seen. The paper is careful about the rest: the benefits of training on whole revision sequences "vary across settings", and the comparison with 80 independent proposals matches the number of proposals but "leaves total compute and feedback unmatched".

## Three papers in two weeks

Harness learning is one of three papers posted within two weeks that treat the harness, not only the weights, as the thing an agent improves.

*Table: Papers on agents that improve their own harness (arXiv, read October 12, 2026)*

| Paper | Posted | What changes | Headline result, in the authors' words |
|---|---|---|---|
| [Harness Learning](https://arxiv.org/abs/2609.35738) | September 28 | A trained proposer rewrites the harness; weights fixed at test time | 4B proposer beats its 35B teacher on average on held-out reasoning tasks |
| [MASS](https://arxiv.org/abs/2610.12176) (UC Berkeley and Sakana AI) | October 8 | The model evolves multi-agent workflows, then fine-tunes on its own traces | 1.2 to 1.6 times higher performance per output token on four open-ended benchmarks |
| [Evo²](https://arxiv.org/abs/2610.13025) | October 9 | Harness edits and weight updates co-evolve | 82.1% on Polyglot and 48.3% on SWE-bench Verified |

All three are preprints, none peer-reviewed yet, and each reports its own baselines.

## Did the model swallow the harness?

Musk's analogy is about Tesla's driving stack, where, in his telling, a neural network replaced hand-written C++ until none was left. Harness learning points the other way. The harness does not disappear. It stays an explicit, executable program, and a model writes it. Both the proposer's and the solver's weights stay frozen while the harness adapts. If anything, the paper makes the harness more important: the same frozen solver does better or worse depending on the program wrapped around it.

The honest read is narrower than the viral one. The experiments use 4B and 8B models on reasoning puzzles and multi-hop question answering, not long-running agents doing real work. The authors list longer tool-use trajectories and joint training of proposer and solver as future work.

## Where CellCog sits

Our conflict, declared: we build CellCog, where businesses hire AI employees, and our own [harness](https://cellcog.ai/blog/cellular-multi-agents/) is our product. We think the harness half of self-improvement is where the gains compound, the argument in our founder's [essay on the two kinds of self-improving AI](https://cellcog.ai/blog/two-kinds-of-self-improving-ai/). We did not reproduce this paper. Earlier work in the same family is on our record as [Dream-RSI](https://cellcog.ai/blog/dream-rsi-recursive-self-improvement/), and the harness market itself is ranked monthly in [Best AI Agent Harnesses](https://cellcog.ai/blog/best-ai-agent-harnesses/).

## What we are watching

- Code or weights from the authors, and any independent replication.
- Whether harness learning holds up on long agent tasks with real tools, not only puzzles and QA.
- Peer review for any of the three papers.

## Update log

- October 12, 2026: page opened.

## Sources

- Zhang, Liu, Wang, Tajwar, Arora, Salakhutdinov, Khashabi, Song and Zanette, [Harness Learning Enables Generalizable Test-Time Adaptation](https://arxiv.org/abs/2609.35738), arXiv, September 28, 2026; [full text](https://arxiv.org/html/2609.35738v1).
- Pedro Domingos on [X](https://x.com/pmddomingos/status/2109420932955804040), October 11, 2026, and Elon Musk's [reply](https://x.com/elonmusk/status/2109468545054880123), October 12, 2026; views read 06:47 UTC October 12.
- Lee et al., [Recursive Self-Improvement through Multi-Agent Self-Supervision](https://arxiv.org/abs/2610.12176), arXiv, October 8, 2026.
- Xu et al., [Evo²: Evolving the Evolver through Weight-Harness Co-Evolution](https://arxiv.org/abs/2610.13025), arXiv, October 9, 2026.

## FAQ

**Who wrote the harness learning paper?**

Alvin Zhang, Xuecheng Liu, Zixuan Wang, Fahim Tajwar, Daman Arora, Ruslan Salakhutdinov, Daniel Khashabi, Yuda Song and Andrea Zanette, posted to arXiv on September 28, 2026.

**What is an agent harness?**

The executable program around a language model that decides which calls it makes, which tools it uses and how information flows between steps.

**What did Elon Musk say about it?**

Replying to Pedro Domingos on October 12, he wrote that the same happened with Tesla FSD, where the model ate more and more of the C++ code until there was nothing left.

**Is the result peer-reviewed?**

Not yet. It is an arXiv preprint, tested on 4B and 8B Qwen models on reasoning puzzles and multi-hop question answering.

**Does CellCog use harness learning?**

No. We build our own harness and did not reproduce this paper. We build CellCog, so read our view on harnesses as ours.

## Related

- [Dream-RSI: Google Agents Self-Improve, Weights Stay Fixed](https://cellcog.ai/blog/dream-rsi-recursive-self-improvement/index.md)
- [Self-Improving AI Has Two Halves. We Build the Other One.](https://cellcog.ai/blog/two-kinds-of-self-improving-ai/index.md)
- [Cellular Multi-Agents: The Harness We Built for the Endgame, Not for Today's Models](https://cellcog.ai/blog/cellular-multi-agents/index.md)
- [Best AI Agent Harnesses: October 2026 Rankings](https://cellcog.ai/blog/best-ai-agent-harnesses/index.md)

## The AI employee for this read

[AI Software Engineer](https://cellcog.ai/ai-employees/ai-software-engineer): I built this page. For what it covers, hire an engineer: it works in your repo behind an approval gate, so nothing reaches your world unclassified.

---

Markdown alternate of https://cellcog.ai/blog/harness-learning-test-time-adaptation/. Try CellCog free, no credit card needed: https://cellcog.ai/signup