# Small editors learn to rewrite agents at test time _Zhang et al. train a 4B proposer to edit frozen-agent harnesses from execution feedback, beating a 35B teacher on held-out Reasoning Gym families._ **Published:** 2026-10-12 **Source:** https://www.startuphub.ai/ai-news/ai-research/2026/small-editors-learn-to-rewrite-agents-at-test-time --- Alvin Zhang and colleagues train a small model to rewrite the program around a frozen solver, and [harness learning](https://arxiv.org/html/2609.35738v1) on arxiv.org frames those edits as meta-learning over executable code. A harness is the glue that decides what the model sees, which tools fire, and how observations loop back. The proposer reads a task description, the current harness and an execution report, then emits a code diff. Training uses reinforcement learning with verifiable rewards, specifically GRPO, where the reward is the mean grader score of the revised harness on a question set. At test time both proposer and solver stay frozen and only the harness changes. The point is adaptation without a weight update. On Reasoning Gym the 4B proposer improved revision quality on families excluded from supervised fine-tuning and RL, and in single-step revision it beat the 35B teacher on average. Trained policies kept improving harnesses over multiple rounds of feedback. The same transfer held for retrieval workflows. A proposer trained only on HotpotQA continued to lift performance on unseen MuSiQue and 2WikiMultihopQA, and in QA even a proposer that had only seen seed harnesses during training generalized to revising already-revised harnesses. Outside this paper, public reasoning leaderboards now stress multi-step and long-context checks like MuSR, SimpleQA, LongBench v2 and MRCRv2, which sharpens why harness-level fixes matter for deployed agents. The timeline matters. The work builds on harness optimization lines from 2023 to 2024, including Shinn et al. on execution-guided reflection and Khattab et al. on DSPy and Yang et al. on SWE-agent, plus the meta-learning formulation of Duan et al. 2016 and Finn et al. 2017 that treats adaptation itself as learnable. Harness-R1, JIT-Agent and TTHE arrived as concurrent harness editors in 2026. What still has to happen is in the paper's limits: dependence on teacher capability for initialization, training the proposer with agentic rollouts rather than static edits, and joint proposer-solver training on longer-horizon tasks. For startups the implication is narrow and practical. If a workflow can be versioned as code and graded automatically, a small editor can compound gains at test time without redeploying the base model, but only where execution feedback is reliable and cheap to collect. --- Original analysis from [startuphub.ai](https://www.startuphub.ai), the #1 AI startup directory. © StartupHub.ai. All rights reserved. You may not republish this article in full without a license. Search engines and AI research tools may crawl and summarize for reference. Bulk reproduction or model training on this content requires a license. See https://www.startuphub.ai/terms.