GenericAgent
A minimal, self-evolving autonomous agent framework built from about 3,000 lines of code and nine atomic tools, giving a language model system-level control of a local computer and growing new skills from every task.
GenericAgent is a minimal, self-evolving autonomous agent framework whose core is about 3,000 lines of code. Through nine atomic tools and an agent loop of roughly 100 lines, it gives a connected large language model system-level control of a local computer, covering the browser, terminal, file system, keyboard and mouse input, screen vision, and Android devices through ADB.
Its design principle is not to preload skills but to evolve them: each time it solves a new task, it crystallizes the steps it took into a reusable skill, so a personal skill tree grows the more the agent is used, starting from the same small seed codebase. A browser-automation component injects into a real browser session so it can preserve existing logins, and the framework works with Claude, Gemini, Kimi, MiniMax and other models across platforms.
The project reports achieving full system control while using substantially fewer tokens than comparable agents, through this smaller, skill-based architecture rather than a large fixed toolset.