SkillEvo: Sustaining Agent Skill Evolution Through Multi-Turn Feedback

Introduction

Agent skills package reusable procedures, prompts, rules, or execution strategies. In many systems, they are either written by humans or generated once by a large language model. Once deployed, however, failures observed during interaction do not necessarily flow back into a reliable mechanism for revising the skill. This makes self-improvement fragile: a system may patch obvious flaws while remaining blind to problems that appear only after context accumulates over several turns.

SkillEvo focuses on this feedback problem. Its central claim is that sustained evolution depends less on whether a system can rewrite a skill, or how many revision rounds it runs, than on whether its evaluation continues to provide a trustworthy “evolution gradient.” In practical terms, the feedback should help identify what failed, expose the underlying weakness, and indicate a useful direction for the next revision.

Key ideas

• From one-shot evaluation to multi-turn feedback. Single-turn question answering can reveal immediate errors, but its signal often weakens after the first round of obvious repairs. SkillEvo treats simulated users not merely as an endpoint for evaluation, but as a source of continuing feedback. Follow-up questions are used to probe what happens after the initial answer.
• Layered discovery of interaction failures. Multi-turn exchanges can test whether a skill preserves consistency, uses earlier information correctly, and avoids creating downstream problems through an earlier response. This broadens evaluation from judging one answer to examining behavior over a sequence of dependent decisions.
• A renewable feedback loop. Each revision consumes the feedback that motivated it, but it also changes the interaction trajectory and can expose new weaknesses. The resulting loop is intended to prevent evolution from stalling once the defects visible in a single exchange have been repaired.
• Active governance rather than passive gating. An end-to-end verification score can reject a degraded candidate, but it generally does not explain or repair the structural cause of degradation. SkillEvo introduces a separate governance layer aimed at correcting factual deterioration and excessive structural growth while constraining the direction of evolution.

Why it matters

The paper highlights an important distinction between a high evaluation score and a useful improvement signal. If evaluation covers only isolated questions, a skill can achieve superficial gains while failing on context accumulation, task handoff, or long-horizon execution. Multi-turn simulation is therefore presented not only as a more demanding test, but also as a mechanism for generating the signals needed by the next revision.

The governance component adds a second systems lesson. Making a skill more capable and keeping it controlled are different objectives. Repeated rewriting may introduce factual errors, redundant instructions, or an increasingly bloated structure. A practical evolution system consequently needs three abilities: discover failures, produce targeted changes, and limit side effects.

SkillEvo frames these responsibilities as two cooperating layers: interaction feedback supplies the direction, while governance constrains the search. The available abstract does not provide specific experimental figures, so the contribution should be read primarily as a framework proposal for multi-turn feedback and controlled skill evolution. Further evaluation will need to establish whether the generated gradients produce stable improvements and whether governance can reduce degradation without suppressing useful exploration.

Source: Hugging Face Daily Papers