Does training editors on real outcomes beat prompting larger models?

The Harness-R1 paper argues that editing an agent's runtime harness (the layer that "constructs context, mediates tools, validates actions, and recovers execution") is better done by a dedicated model trained on outcomes than by a prompted frontier model. Its discussion states the claim directly: "Training a dedicated engineer beats prompting a larger model." The abstract reports the headline result for the trained engineer: vanilla Qwen3.5-9B success rises from 44.3% to 53.6% (+9.3 points) across WebShop, ALFWorld, and DBBench. The authors say prompted frontier editors produce gains that are "unstable and sometimes even lower reward."

The mechanism is where the outcome signal comes from. A separate 9B engineer turns batches of target-agent failures into "validated executable patches," and fresh same-batch reruns of the frozen target supply the reward. Training therefore updates only the engineer, initialized by cold-start supervised fine-tuning and then trained online with group-relative policy optimization. The paper's explanation for the prompted-editor failure is not size but feedback. Fixed frontier models "optimize for plausibility, emitting syntactically valid and reasonable-looking edits," yet "never rerun the target," so they cannot tell whether an edit raised task success. The reasoning is that an edit "intervenes directly in a running executable system," so it has to be "precise, verifiable, and genuinely beneficial," which text that merely looks right does not guarantee.

This sits in tension with Do stronger models always evolve harnesses better?, where a 9B evolver induced gains comparable to a much stronger one. The two agree that raw editor size is not the lever, but they point to different levers: that note points to investing in the agent that uses the harness, while this paper points to training the editor against realized outcomes. It also gives an outcome-trained alternative to the external anchors in Does constraining edits make skill learning more stable?, where a held-out gate and a rejected-edit buffer stop locally plausible edits from accumulating. Both treat plausibility as an insufficient test for a self-edit and put a check outside the editor's own judgment. The reward described here comes from rerunning the same batch of failures, which raises the concern in Do harness edits learn reusable strategies or memorize task fixes?: a same-batch reward cannot by itself separate a memorized fix from a reusable one.

The excerpt leaves several things open. It does not name the prompted frontier models, say how much larger they are, or give their numbers. The 9.3-point figure is the gain over the unedited target, not over those prompted baselines, and the introduction evidence the discussion cites ("as shown in the introduction") is not in the excerpt. It also does not say whether patches transfer to held-out tasks or a different target agent, or what the rerun loop costs. At the strength the evidence allows, the claim is that an outcome-conditioned editing signal beats prompting for harness edits in these three environments with a 9B target. The excerpt cannot say whether the gain comes from the training or from the rerun feedback alone.