Self-Learning Web Agents: SkillWeaver Enables Continuous Improvement Through API Synthesis

Interacting with websites requires a variety of skills, from navigation and form input to content interpretation. Autonomous web agents designed to automate these tasks face the challenge of mastering these complex processes. A new approach called SkillWeaver promises a solution by enabling agents to independently improve their skills and store them as reusable APIs.

SkillWeaver is based on a skill-centric framework. Confronted with a new website, the agent analyzes the structure and possible interactions. In doing so, it independently discovers recurring patterns and abstracts them as individual skills. The agent then practices executing these skills and distills the collected experience into robust APIs. These APIs act as building blocks for more complex tasks and can be flexibly combined.

Through iterative exploration and the continuous addition of new skills, the agent constantly expands its API catalog. This growing collection of lightweight and easily integrable APIs significantly improves the agent's capabilities. It not only learns to operate individual websites more efficiently, but can also transfer its knowledge to new, unknown sites.

The effectiveness of SkillWeaver has been tested in experiments on the WebArena platform and with real-world websites. The results show a significant increase in the agents' success rate. On WebArena, a relative improvement of 31.8% was achieved, and even 39.8% on real-world websites. Particularly noteworthy is the possibility of skill transfer: APIs synthesized by high-performing agents can be passed on to less capable agents. This led to an improvement in the success rate of up to 54.3% on WebArena.

The ability to abstract different website interactions into APIs and exchange them between agents opens up new possibilities for the development of autonomous web agents. SkillWeaver paves the way for scalable, robust, and adaptable agents that are able to independently adapt to the ever-growing complexity of the internet.

The development of SkillWeaver is an important step towards more powerful and flexible web agents. The ability to learn from experience and store what has been learned in reusable modules enables continuous improvement and adaptation to new challenges. The possibility of skill transfer between agents accelerates the learning process and promotes the development of a shared knowledge base. Future research will focus on optimizing skill discovery, improving API robustness, and developing efficient strategies for skill transfer.

Mindverse, a German company specializing in AI-powered content creation, image generation, and research, is following developments in the field of autonomous web agents with great interest. As a provider of customized AI solutions, including chatbots, voicebots, AI search engines, and knowledge systems, Mindverse recognizes the potential of technologies like SkillWeaver for optimizing and automating web interactions. The integration of such innovative approaches into its own products and services is a central concern of Mindverse.

Bibliography: Zheng, B., Fatemi, M. Y., Jin, X., Wang, Z. Z., Gandhi, A., Song, Y., Gu, Y., Srinivasa, J., Liu, G., Neubig, G., & Su, Y. (2025). SkillWeaver: Web Agents can Self-Improve by Discovering and Honing Skills. arXiv preprint arXiv:2504.07079. Li, Y., Yu, T., Zeng, A., Gu, S. S., Song, Y., Liu, S., ... & Han, J. (2024). Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2405.20309. Yao, S., Yu, D., Song, J., Dong, Q., Poole, B., & Yu, T. (2024). React: Synergizing reasoning and acting in language models. OpenReview.