Does constraining edits make skill learning more stable?

The prevailing self-improvement recipe lets an agent rewrite its own instructions freely from feedback. SkillOpt's ablations argue this is exactly wrong: bounded textual learning outperforms uncontrolled rewriting. A textual learning-rate budget limits how far one skill version may move from the previous one; a held-out gate prevents harmful proposals from accumulating; a rejected-edit buffer retains failed edits as explicit negative feedback so the optimizer does not re-propose them; and an epoch-wise slow/meta update preserves long-horizon regularities without bloating the deployed skill.

This matters because uncontrolled self-revision has a characteristic failure: each edit looks locally plausible, but unchecked accumulation drifts the skill toward instance-specific overfitting or incoherent sprawl. The constraints are not bureaucratic overhead — they are what convert noisy self-edits into a stable optimization trajectory. The rejected-edit buffer is the subtle piece: a failed edit is usually discarded, but as retained negative feedback it carries information about what not to do, much as hard negatives sharpen contrastive learning.

The counterpoint is that bounding edits trades adaptability for stability — too tight a learning rate could prevent the skill from escaping a poor starting point. But SkillOpt's per-benchmark case studies show the learned skills stay compact, inspectable, and procedural rather than instance-specific, suggesting the bound is doing its intended job. Therefore the pattern generalizes to any self-editing system: durable self-improvement comes from controlled, validated, memory-of-failures editing — not from giving the model maximal freedom to rewrite itself.