GenericAgent is a minimal, self-evolving autonomous agent framework that starts from a tiny seed of code and grows its own skill tree from experience. It has more than 14,300 stars on GitHub and was actively updated in 2026.
The whole system grows from roughly 3,000 lines of seed code, nine atomic tools and an agent loop of about 100 lines. Whenever GenericAgent solves a task, it distills the successful procedure into a reusable skill, so repeat work gets faster and cheaper — the project reports up to six times lower token consumption on tasks it has learned. Through its atomic tools it can control the full system, including browser automation and desktop operation, and it connects to leading models such as Claude and Gemini.
The design philosophy is radical minimalism: rather than shipping a giant framework of predefined workflows, GenericAgent trusts the loop of acting, verifying and recording what worked. Its tools span the surfaces an autonomous operator needs — the filesystem, a shell, web browsing and direct desktop control — and the skill tree it grows is stored in inspectable form, so users can audit, edit or share the procedures it has learned. A public skill hub and a technical report accompany the code, and the project is aimed at people who want to understand every line of the agent they run.
GenericAgent is free and open source under the MIT license. Your only cost is the LLM API usage of the model you connect to it.
Agent researchers, automation hobbyists and framework minimalists who want an agent that learns from its own work instead of starting from zero every time.