SkillOpt-Lite - Better and Faster Agent Self-evolution via One Line of Vibe

Sep 21, 2026

Uploaded by Oliver

SkillOpt-Lite proposes a minimal viable pipeline for autonomously optimizing agent skills without complex multi-stage components. The method formalizes skill optimization as Zeroth-Order optimization, translating classical ideas like trust regions and central differences into trajectory-based, semantically rich execution traces for interpretable debugging feedback. Grounded in PAC learning and Claude Code philosophy, it applies file-system trajectory exploration, consensus attribute mining, and independent validation gating. SkillOpt-Lite accelerates convergence and improves LiveMath and SpreadsheetBench, and supports production coding agents such as VSCode Copilot.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages16

Reading Time40 min

AccessFree to download

Introduction

Minimal viable pipeline for skill optimization

Zeroth-Order (ZO) formalization

From trajectories to interpretable debugging feedback

Design principles for convergence and generalization

Consensus Mining across Trajectories

Independent validation gating

Evaluation results

LiveMath improvements

SpreadsheetBench and HarnessOpt

arXiv :2607 .0345 1v 1 [ cs . SE] 3 Jul 2026 SkillOpt-Lite: Better and Faster Agent Self-evolution via One Line of Vibe Yifei Shen 1 Bo Li 1 ,2 Xinjie Zhang3 1 LMMs-Lab 2 NTU MMLab 3 Microsoft While skill optimization for autonomous agents has gained traction, existing methods rely on complex pipelines. This leaves a fundamental question unaddressed: What constitutes a minimal viable pipeline for skill optimization, where every component is justified by theory or empirical necessity? We formalize skill optimization via Zeroth-Order (ZO) optimization, mapping classical counterparts (central difference, trust regions) to recent literature. Noting that unlike blind numerical perturbations in classical ZO, skill trajectories serve as interpretable debugging feedback. Grounded in Claude Code philosophy and PAC learning, we establish three principles for convergence and generalization: file-system-based trajectory exploration, consensus attribute mining, and independent validation gating. Eliminating redundancies, we propose SkillOpt-Lite. It accelerates convergence and outperforms full SkillOpt: improving LiveMath by +8.8 points on GPT-5.5 and +25.4 points on GPT-5.4-nano, allowing the nano model to surpass standard GPT-5.4 optimized by SkillOpt. Finally, we integrate our framework into production coding agents like VSCode Copilot, enabling developers to evolve agent skills via one line of vibe. Because our framework treats all agent components simply as standard editable code, this minimal pipeline naturally generalizes to full harness optimization (HarnessOpt) . On SpreadsheetBench, HarnessOpt enables GPT-5.4-nano to achieve 0.7758 accuracy, outperforming the larger GPT-5.5 running standard pipelines (0.7620). Code is available at [https://github.com/EvolvingLMMs-Lab/SkillOpt-Lite](https://github.com/EvolvingLMMs-Lab/SkillOpt-Lite). 1 Introduction AI agents have transitioned from research prototypes into standard production utilities [1–3] . Practical deployment reveals that an agent’s real-world capability is determined not solely by its base Large Language Model (LLM), but by the interaction among the base model, its operational scaffolding (harness), and its domain heuristics (skills) . Because foundation models remain frozen during inference, agent engineering frequently reduces to refining these skill documents—a process where subtle textual variations routinely drive non-linear disparities in downstream task performance [4–8] . To manage this non-linear sensitivity, automated skill optimization has evolved from open-loop self-reflections [9, 10] into complex algorithmic pipelines [8, 11–13] . Among existing literature, SkillOpt [4] formalizes text revision through continuous optimization analogies, implementing minibatch tree merging, textual learning-rate schedules, and rejected-edit buffers. Yet, despite their empirical gains, existing frameworks exhibit a growing tendency toward architectural complexity. They incorporate intricate mechanics from classical optimization without addressing a fundamental question: What constitutes a minimal viable pipeline for skill optimization, where every component is justified by theoretical or empirical necessity? SearchQA 73.5 73.4 DocVQA Spreadsheet OfficeQA 38.6 40.0 ALFWorld 63.5 LiveMath Init skill SkillOpt SkillOpt-Lite (a) Macro-level average performance profiling. Best val_hard so far Best val_hard so far 0.6 0.5 0.4 0.3 0.2 0.8 0.7 0.6 0.5 0.4 LiveMath - GPT-5 .4-nano Spreadsheet - GPT-5.5 1 2 3 4 5 6 7 8 9 10 Batch 0.8 0.7 0.6 0.5 0.4 1.000 0.975 0.950 0.925 0.900 0.875 0.850 0.825 0.800 LiveMath - GPT-5 .5 DocVQA - GPT-5 .4 | | | | | | | | | | | | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |

FAQs

What problem does SkillOpt-Lite address in agent skill optimization?

It targets the lack of a clearly justified minimal pipeline for skill optimization, since existing methods often rely on complex components without theoretical or empirical necessity.

How does the paper connect skill optimization to Zeroth-Order optimization?

It maps prior reflection paradigms to ZO optimization counterparts, while highlighting a key divergence: agent rollouts provide semantic, observable trajectories rather than blind perturbations of unobservable states.

What are the core principles used by SkillOpt-Lite?

It uses file-system-based trajectory exploration, consensus attribute mining across trajectories, and independent validation gating to support convergence and generalization.

Sep 21, 2026

Uploaded by Oliver