arXivPaperNeeds Review

SkillProx: Self-Evolving Agent Skills via Proximal Textual Gradient Descent

SkillProx evolves reusable agent instructions through measured retries, regression rollback, and utility-based pruning, reporting a 3-point accuracy gain over its strongest baseline.

arXiv · Aug 7, 2026

Open Source Open Markdown Open JSON

Source Summary

SkillProx treats agent skills as editable text rather than model weights. Its **forward stage** reruns diagnosis-driven edits on the same task batch, rolls back regressions, and returns measured outcomes; its **backward stage** audits individual knowledge units for consolidation, demotion, or removal.

Practical Implication

For recurring coding-agent work, the useful pattern is closed-loop skill maintenance: test each instruction change, preserve the prior version when results worsen, and periodically remove guidance that contributes no measured utility. This makes accumulated procedures more auditable than an ever-growing prompt file.

Agent-Ready Context

SkillProx treats agent skills as editable text rather than model weights. Its **forward stage** reruns diagnosis-driven edits on the same task batch, rolls back regressions, and returns measured outcomes; its **backward stage** audits individual knowledge units for consolidation, demotion, or removal.

For recurring coding-agent work, the useful pattern is closed-loop skill maintenance: test each instruction change, preserve the prior version when results worsen, and periodically remove guidance that contributes no measured utility. This makes accumulated procedures more auditable than an ever-growing prompt file.

Across multiple backbones and in- and out-of-distribution benchmarks, SkillProx reports an average **3.0 percentage-point** accuracy gain over the strongest gradient-based baseline. The abstract does not identify the tasks, absolute scores, audit cost, or whether the approach transfers to live software repositories.

Connected Context · Feed7 Judgment

This operationalizes skill governance as reversible, measured maintenance: edit instructions, rerun the same tasks, roll back regressions, and periodically prune low-utility knowledge. It directly addresses the danger of endlessly accumulating procedures, but does not establish repository-scale practicality because task details, absolute results, audit cost, and live coding transfer are absent.

The Regression Tax: Decomposing Why Skills Help and Hurt LLM AgentsSkillProx turns the Regression Tax finding into a maintenance mechanism by explicitly detecting regressions, preserving prior versions, and auditing individual knowledge units instead of judging only aggregate gains.Don't Ship Skills Without Evals — Philipp Schmid, Google DeepMindThe repeated with-and-without regression testing advocated here is a prerequisite for SkillProx’s rollback and pruning decisions; without harness-specific trials, its textual edits cannot be assigned measured utility reliably.EveryInc/compound-engineering-pluginCompound Engineering supplies the kind of persisted workflows and solved-problem notes SkillProx could maintain, while SkillProx adds a missing discipline for testing, reverting, consolidating, or removing that accumulated guidance.WTF Is the Context Layer? The Missing Infrastructure for Production Agents — Prukalpa SankarAtlan’s shared, versioned context broadens the storage and portability layer for skills, while SkillProx contributes a candidate method for deciding which versioned instructions should be promoted, demoted, or removed.

Context Map

agentcoding#skills#agent-memory #agent-evals

Uncertainty

Across multiple backbones and in- and out-of-distribution benchmarks, SkillProx reports an average **3.0 percentage-point** accuracy gain over the strongest gradient-based baseline. The abstract does not identify the tasks, absolute scores, audit cost, or whether the approach transfers to live software repositories.