# SkillWeaver: Web Agents can Self-Improve by Discovering and Honing Skills

**Authors:** Boyuan Zheng, Michael Y. Fatemi, Xiaolong Jin, Zora Zhiruo Wang, Apurva Gandhi, Yueqi Song, Yu Gu, Jayanth Srinivasa, Gaowen Liu, Graham Neubig, Yu Su  
**arXiv:** [2504.07079](https://arxiv.org/abs/2504.07079)  
**Date:** 2025-04-09  
**Summary page:** https://www.summarizepaper.com/en/paper/2504.07079v1/

## Key points

- Humans have evolved self-improvement mechanisms through environment exploration, hierarchical abstraction of experiences, and collaborative construction of skill repertoires
- Autonomous web agents struggle with procedural knowledge abstraction, refining skills, and skill composition
- SkillWeaver framework enables agents to synthesize reusable skills as APIs through three stages:
- Skill Proposal: Identifying novel skills based on observations and available APIs
- Skill Synthesis: Generating successful trajectories from proposed skills to synthesize APIs
- Skill Honing: Testing synthesized APIs with automatically generated test cases for robustness
- Experiments show significant success rate improvements on WebArena and real-world websites using SkillWeaver
- Stronger agents can enhance weaker ones through transferable skills synthesized by SkillWeaver
- SkillWeaver enables autonomous self-improvement in navigating complex online environments by building conceptual maps, accumulating procedural knowledge as reusable skills, composing simple skills into complex routines, and enhancing decision-making processes without extensive training data or external supervision

## Summary

In complex environments, humans have evolved sophisticated self-improvement mechanisms through environment exploration, hierarchical abstraction of experiences into reusable skills, and collaborative construction of a skill repertoire. However, autonomous web agents still struggle with procedural knowledge abstraction, refining skills, and skill composition. To address this challenge, SkillWeaver is introduced as a skill-centric framework enabling agents to autonomously synthesize reusable skills as APIs. The framework consists of three stages: 

1. Skill Proposal: The agent identifies novel skills based on observations and available APIs in the skill library. 2. Skill Synthesis: Successful trajectories generated by executing proposed skills are used to synthesize APIs. 3. Skill Honing: Synthesized APIs undergo testing with automatically generated test cases for robustness. Through iterative exploration and practice on new websites, the agent builds a library of lightweight, plug-and-play APIs that enhance its capabilities significantly. Experiments on WebArena and real-world websites demonstrate the efficacy of SkillWeaver, with relative success rate improvements of 31.8% and 39.8%, respectively. Stronger agents can also enhance weaker ones through transferable skills synthesized by SkillWeaver, resulting in improvements of up to 54.3% on WebArena. By honing diverse website interactions into APIs that can be seamlessly shared among web agents, SkillWeaver showcases the effectiveness of autonomous self-improvement in navigating complex online environments. This framework enables agents to build conceptual maps of website environments, accumulate procedural knowledge as reusable skills, compose simple skills into complex routines, and enhance decision-making processes through learned skills without the need for extensive training data or external supervision.