GenericAgent is a minimal, self-evolving autonomous agent framework built on approximately 3,000 lines of core Python code. The repository demonstrates a novel approach to agent design by starting with a lean seed codebase and allowing the system to grow its capabilities organically through task execution. Rather than preloading extensive skill libraries, GenericAgent automatically crystallizes execution paths into reusable skills that accumulate over time, forming a personal skill tree that expands with each new task the agent solves.
The framework grants large language models system-level control over local computers through nine atomic tools and a roughly 100-line agent loop. These tools provide comprehensive system access including browser automation, terminal commands, filesystem operations, keyboard and mouse input, screen vision capabilities, and mobile device control via ADB. A distinctive feature is the use of TMWebdriver, which injects directly into real browser instances while preserving login sessions, enabling the agent to handle real-world challenges like CAPTCHA verification without relying on headless browser workarounds.
GenericAgent achieves significant token efficiency, consuming less than 30,000 tokens in context windows compared to other agents that require 200,000 to 1,000,000 tokens. This efficiency translates to reduced hallucinations, higher task success rates, and substantially lower operational costs. The framework supports multiple major language models including Claude, Gemini, Kimi, and MiniMax, with cross-platform compatibility across Windows, macOS, and Linux.
The repository's self-bootstrap proof is notable: the entire codebase, from initial Git setup through all commit messages, was completed autonomously by GenericAgent itself without human terminal interaction. This demonstrates the framework's practical viability for autonomous task execution. The demonstration gallery showcases diverse capabilities including real-browser CAPTCHA survival, autonomous web exploration, food delivery ordering, quantitative stock screening, expense tracking, and batch messaging across multiple platforms.
Installation is deliberately tiered to minimize dependencies. The core agent requires only the requests library plus four lightweight packages for the local server supporting TMWebdriver. Optional UI extras including Streamlit and terminal UI components can be installed separately. The framework provides multiple frontend interfaces: a terminal UI built on prompt_toolkit and rich, a Streamlit web interface, and instant messaging integrations with Telegram, Discord, and Lark.
Advanced capabilities are unlocked through natural language instructions rather than configuration files or documentation reading. Users can enable web automation, OCR, vision capabilities, and computer use by simply instructing the agent to set up these features. The agent then consults pre-installed standard operating procedures, installs missing dependencies, adapts to the operating system, and persists configurations to memory.
According to GitGenius activity tracking, the repository shows median issue and pull request response latency of 3.0 hours across 87 tracked items, with mean latency of 62.2 hours. Primary contributor lsdefine has logged 44 events, with secondary contributors desmonna and szdtzpj each contributing 6 events. The repository connects to related projects through overlapping contributors, including alirezarezvani/claude-skills, huggingface/transformers, and huggingface/datasets. The framework is classified across multiple AI domains including agent systems, LLM integration, tool usage, memory management, and task automation, reflecting its comprehensive approach to autonomous agent design.
