GenericAgent Grows Its Own Skill Tree from 3K Lines of Code
This one is wild. GenericAgent is a self-evolving agent framework with just 3,000 lines of core code and a 100-line agent loop. Give it a task, it solves it, then automatically crystallizes the execution path into a reusable skill. Over time, it grows a personal skill tree that makes it faster and cheaper at everything it's done before.
Nine atomic tools give it full system-level control: browser, terminal, filesystem, keyboard and mouse input, screen vision, and even mobile devices via ADB. The claim on the repo is that everything in it — including installing Git and writing every commit message — was done autonomously by GenericAgent. The author never opened a terminal.
The token efficiency angle is what makes this more than a demo project. GenericAgent claims 6x less token consumption than baseline agents on the same tasks, because learned skills shortcut the reasoning chain. Instead of figuring out how to do something from scratch every time, it reuses crystallized execution paths. This is the same insight behind projects like Memento-Skills and EvoSkill, but GenericAgent packages it into something minimal enough to actually run.
It supports Claude, Gemini, Kimi, MiniMax, and other major models. The five-tier memory system goes from meta-rules down to session archives, giving the agent both persistent knowledge and ephemeral working memory.
Trending at 400+ stars per day on GitHub with 1.9K total stars. The self-evolving agent pattern is clearly resonating. The question is whether skill trees that grow organically can match hand-crafted agent workflows — but the 6x token reduction suggests they might already be winning on efficiency.
https://github.com/lsdefine/GenericAgent
← Back to all articles
Nine atomic tools give it full system-level control: browser, terminal, filesystem, keyboard and mouse input, screen vision, and even mobile devices via ADB. The claim on the repo is that everything in it — including installing Git and writing every commit message — was done autonomously by GenericAgent. The author never opened a terminal.
The token efficiency angle is what makes this more than a demo project. GenericAgent claims 6x less token consumption than baseline agents on the same tasks, because learned skills shortcut the reasoning chain. Instead of figuring out how to do something from scratch every time, it reuses crystallized execution paths. This is the same insight behind projects like Memento-Skills and EvoSkill, but GenericAgent packages it into something minimal enough to actually run.
It supports Claude, Gemini, Kimi, MiniMax, and other major models. The five-tier memory system goes from meta-rules down to session archives, giving the agent both persistent knowledge and ephemeral working memory.
Trending at 400+ stars per day on GitHub with 1.9K total stars. The self-evolving agent pattern is clearly resonating. The question is whether skill trees that grow organically can match hand-crafted agent workflows — but the 6x token reduction suggests they might already be winning on efficiency.
https://github.com/lsdefine/GenericAgent
Comments