When an AI rewrites its own instructions, does a bigger editor help — or does seeing whether the fix worked matter more?

This explores why, when a model edits something like an agent's harness, a prompt, or a skill file, what tells it whether an edit worked matters more than how big the editor is. The sharpest evidence is a head-to-head. A 9B model trained with reinforcement learning on whether its patches actually worked lifted a frozen agent by 9.3 points across three tasks. Prompted frontier models gave unstable or lower gains Does training editors on real outcomes beat prompting larger models?. The trained editor reran each patch and saw what happened. The prompted models were only optimizing for plausibility, meaning an edit that reads like a good fix.

Plausibility is a feedback signal too, just a bad one. In one production case, a prompt mutation raised a rationale-alignment pass rate from 23.1% to 80.0% by adopting the judge's preferred vocabulary, while defect-identification precision didn't move Can prompt optimization accidentally teach judges to reward the wrong signals?. The prompt learned to sound right rather than be right. Nothing in that loop rewards being correct, so extra model capacity has nothing true to aim at. A bigger editor would probably just find the shortcut faster.

The corpus also suggests the shape of the signal matters, not just whether it's an outcome. A survey of 259 works finds verification loops guide repair only when the feedback points at a scope the system's available actions can actually fix When does verification feedback actually guide targeted artifact repair?. SkillOpt gets more stable skill learning by constraining the editor rather than enlarging it. It caps how much can change per step, checks edits against held-out validation, and keeps rejected edits as negative feedback Does constraining edits make skill learning more stable?. Step-level critique during training does something similar for diversity, preventing solutions from narrowing too early Do critique models improve diversity during training itself?.

Size isn't a reliable lever in general. For synthetic data generation, models around 500M parameters produce more unique outputs per sample than larger ones, because larger models pile probability onto their preferred answers Why aren't bigger models better for generating diverse outputs?. That's a different task, but it makes the same point: bigger doesn't win on the axis you care about.

Only one note here tests editor size against feedback source directly, so the rest is supporting evidence, not proof. The self-evolving harness Ouroboros posts state-of-the-art scores by editing itself through reviewed commits, but the paper has no ablations to separate the editing from the review signal Does self-editing through reviewed commits improve agent performance?. Whether the review step or the editor does the heavy lifting there is still open.