finds.dev← search

// the find

lsdefine/GenericAgent

★ 7,986 · Python · MIT · updated Apr 2026

Self-evolving agent: grows skill tree from 3.3K-line seed, achieving full system control with 6x less token consumption

GenericAgent is a minimal autonomous agent framework (~3K lines) that gives LLMs full system control via 9 atomic tools (browser, terminal, filesystem, ADB, vision) and a self-evolving skill tree that persists reusable workflows across sessions. It targets developers who want a lightweight local agent without the complexity of frameworks like LangChain or AutoGen, and is primarily aimed at Chinese-market integrations given the heavy WeChat/Alipay/DingTalk tooling.

- The skill crystallization mechanism is genuinely clever: instead of re-prompting from scratch every run, solved tasks get serialized into reusable skill files, and the layered L0-L4 memory system keeps context windows small (~30K) by scoping what gets loaded per session.

- Minimal dependency surface — core runs with just requests, beautifulsoup4, bottle, and websocket-server. Browser injection reuses an existing Chrome session with cookies intact rather than spinning up headless Playwright, which actually matters for sites with MFA or bot detection.

- The self-bootstrap claim (repo was committed entirely by the agent) is verifiable in git history and is the most honest demo you can give for an agentic framework.

- Multiple chat frontends (Telegram, WeChat, DingTalk, Feishu, WeCom) with a clean shared abstraction in chatapp_common.py, making it practical as a personal assistant rather than a demo toy.

- Security model is essentially nonexistent: `code_run` executes arbitrary Python with no sandboxing, and the agent has full system access. Running this against any untrusted input or a compromised LLM provider is a direct path to RCE on your machine.

- Skill tree reliability degrades silently — if the LLM writes a flawed skill on first execution, that bad skill gets recalled forever. There's no skill validation, versioning, or rollback mechanism described in the codebase.

- The project is heavily oriented toward Chinese platforms and workflow patterns (WeChat, Alipay, ADB for Chinese apps), and the English documentation is clearly a translation. International users will hit friction quickly with untranslated skills and Chinese-only community support.

- No test suite whatsoever. For a framework that executes code autonomously and modifies your filesystem, the complete absence of automated tests is a real adoption risk in anything beyond personal use.

View on GitHub → Homepage ↗

// want more like this?

We dig through GitHub every week and send a few repos picked for what you actually care about — each with an honest take like this one.

Get finds in your inbox → Search again →