GenericAgent is an open-source, self-evolving autonomous agent built in roughly 3,000 lines of Python that gives large language models system-level control of a local computer. It positions itself between bare Claude Computer Use and heavier frameworks like OpenHands — a minimal but capable agent that browses the web, runs shell commands, reads and edits files, and orchestrates multi-step tasks through a clean tool-calling loop backed by any OpenAI-compatible endpoint.
What sets GenericAgent apart is its skill crystallization pattern. Instead of re-exploring the same problem on every run, the agent saves the successful solution path as a reusable skill inside a growing skill tree. A task that costs thousands of tokens the first time can be replayed from the crystallized skill for a fraction of the cost, and the tree itself becomes a durable memory of what the agent has learned to do on the host machine. This makes GenericAgent a practical choice for developers who want a local computer agent with persistent, inspectable capabilities rather than one-shot chat sessions.
The project is MIT-licensed and deliberately small — the whole loop, skill store, and tool definitions fit in a repository a solo engineer can read in an afternoon. That makes it a strong base for teams that want to fork an agent, wire in their own tools, and ship an internal automation without committing to a larger framework's abstractions. It is a good fit for self-hosted AI workflows, developer productivity tinkering, and research on long-horizon agent behavior.
