Automatically Learning Skills for Coding Agents
Shangyin Tan (University of California, Berkeley), Lakshya A Agrawal (UC Berkeley), Rohit Sandadi (University of California, Berkeley), Dan Klein (University of California, Berkeley), Koushik Sen (UC Berkeley), Alexandros G. Dimakis (UC Berkeley), Matei Zaharia (UC Berkeley)
System Optimization & Efficiency Architectural Patterns & Composition
Summary
A fully automated pipeline that learns repository-specific skills from synthetic tasks and evolutionary optimization, boosting coding agent performance without fine-tuning.
Description
Coding agents powered by large language models can solve a wide range of software engineering tasks, yet they often struggle on unfamiliar repositories whose conventions, testing patterns, and project structure differ from their training data. We introduce gskill, a fully automated pipeline that learns repository-specific skills, concise natural-language documents that capture the knowledge an agent needs to work effectively within a given codebase. gskill combines two components: (1) SWE-smith, a data-generation pipeline that creates diverse, verifiable software engineering tasks from any GitHub repository, and (2) optimize_anything, an evolutionary optimization loop that iteratively refines skill documents using LLM-generated feedback. Skills learned by gskill on a lightweight agent (Mini-SWE-Agent with GPT-5-mini) transfer directly to stronger agents: on the bleve repository, Claude Code with Claude Haiku 4.5 improves from 79.3% to 100.0% pass rate while also reducing average task duration from 173s to 130s. On jinja, Claude Haiku 4.5 improves from 93.9% to 98.5% while Claude Sonnet 4.5 maintains a perfect 100%. Our results demonstrate that automatically learned, repository-specific skills provide a lightweight and transferable mechanism for improving coding agent performance without model fine-tuning. gskill is publicly available at github.com/gepa-ai/gepa/tree/main/src/gepa/gskill.
