lsdefine/GenericAgent

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

What it solves

GenericAgent is a minimal autonomous agent framework designed to give LLMs system-level control over a local computer. It addresses the problem of bloated agent frameworks by providing a tiny core that allows an agent to autonomously learn and evolve its own capabilities (skills) through experience rather than relying on pre-loaded, static toolsets.

How it works

The framework uses a minimal seed of approximately 3,000 lines of code and a ~100-line execution loop. It provides 9 atomic tools (such as code execution, file manipulation, and web scanning) that allow the agent to interact with the browser, terminal, filesystem, and mobile devices via ADB.

Crucially, it employs a Self-Evolution Mechanism: when the agent solves a new task, it crystallizes the execution path into a reusable "Skill" stored in a layered memory system (L0-L4). This allows the agent to recall and execute complex workflows instantly in the future, reducing token consumption and hallucinations.

Who it’s for

It is intended for developers and power users who want a lightweight, self-evolving AI agent capable of system-level automation across different operating systems and real-browser interaction.

Highlights

  • Self-Evolving Architecture: Automatically turns successful task completions into permanent, reusable skills.
  • Minimalist Core: Extremely small codebase (~3K lines) with low deployment overhead.
  • TMIWebdriver: Uses a real, persistent browser session to preserve logins and bypass bot detection (passing SannySoft and reCAPTCHA v3 tests).
  • Token Efficient: Operates with a context window under 30K tokens, significantly lower than many competing agents.
  • System-Level Control: Direct access to terminal, keyboard/mouse, and mobile devices (ADB).

Related

  • Project
    AtomicBot-ai/atomic-agent

    A local-first, open-source AI agent that runs a control loop on your machine, can operate a browser, edit files, run approved commands, and keeps persistent memory — designed to work efficiently with small quantized models via llama.cpp.

  • Project
    modelscope/ms-agent

    MS‑Agent is an open‑source Python framework that lets you build, run, and manage autonomous AI agents. It provides a customizable harness for planning, tool execution, permission checks, and feedback, with built‑in context‑compression, long‑term memory, and DAG‑style multi‑agent workflows. Agents can be used via a Web UI, terminal UI, CLI, or embedded SDK, and the ecosystem includes reusable “skills”, multimodal input, and ready‑made applications for research, code generation, video creation, and finance.

  • Project
    ANative-Lab/EvoAgentX

    An open-source framework for automatically building, evaluating, and evolving LLM-based agentic workflows through iterative feedback loops.

  • Project
    wanxingai/LightAgent

    LightAgent is an Apache‑2.0‑licensed Python framework for building lightweight, production‑grade LLM agents. It offers built‑in memory (via mem0), tool integration, tree‑of‑thought reasoning, multi‑agent routing (LightSwarm), deterministic workflow orchestration (LightFlow), optional dynamic DAGs, OpenAI‑compatible streaming APIs, structured tracing, guardrails, and human‑approval hooks. Install with `pip install lightagent` and start with a single `LightAgent(...).run(query)` call.

  • Project
    YeQing17-2026/OmniAgent

    An open-source self-evolving agent framework that dynamically updates its skills, memory, and brain model through interaction while employing a multi-layer security scaffold.