GenericAgent is a groundbreaking, self-evolving autonomous agent framework designed to provide LLMs with system-level control over a local computer. Its core strength lies in its simplicity and efficiency, achieved through a minimal codebase of approximately 3,000 lines of code and a streamlined architecture. The project's primary purpose is to empower users with a versatile and adaptable agent that learns and improves over time, eliminating the need for pre-defined skills and complex configurations.
The core functionality of GenericAgent revolves around its self-evolving mechanism. Instead of relying on a pre-loaded set of skills, the agent learns and adapts by crystallizing successful execution paths into reusable skills. When presented with a new task, the agent autonomously explores solutions, which may involve installing dependencies, writing scripts, and debugging. Once a solution is found and verified, the execution path is transformed into a skill and stored in the agent's memory. This skill is then readily available for future use, allowing the agent to perform similar tasks with a single command. This self-evolution process results in a personalized skill tree that grows with each use, making the agent increasingly capable and tailored to the user's specific needs.
The framework boasts several key features that contribute to its effectiveness. It offers a minimal architecture, reducing complexity and deployment overhead. The agent's core loop, responsible for task execution, is only about 100 lines of code. GenericAgent provides strong execution capabilities, enabling it to interact with a real browser, preserving login sessions, and directly controlling the system through nine atomic tools. It also exhibits high compatibility, supporting various LLM providers like Claude, Gemini, Kimi, and MiniMax, and is cross-platform compatible. Furthermore, the agent is token-efficient, utilizing a context window of less than 30,000 tokens, a significant advantage compared to other agents that consume significantly more resources. This efficiency is achieved through a layered memory system that ensures the right knowledge is always within scope, leading to fewer hallucinations, higher success rates, and reduced costs.
The agent's workflow is centered around a layered memory system and a minimal toolset. The layered memory system comprises several layers: L0 (Meta Rules), L1 (Insight Index), L2 (Global Facts), L3 (Task Skills/SOPs), and L4 (Session Archive). These layers work together to store and retrieve information efficiently, enabling the agent to learn from past experiences and adapt to new situations. The minimal toolset consists of nine atomic tools that provide the foundational capabilities for interacting with the external environment. These tools include code execution, file reading and writing, file modification, web content perception, browser control, and user interaction. Additionally, the agent can dynamically create new tools through the `code_run` tool, allowing it to install packages, write scripts, and control hardware at runtime.
GenericAgent's self-evolution mechanism is a key differentiator. The agent's ability to learn and adapt over time makes it a powerful tool for automating complex tasks. The agent's capabilities grow with each use, forming a personalized skill tree that is unique to the user. This self-learning approach allows the agent to become increasingly efficient and effective over time, making it a valuable asset for a wide range of applications. The repository also provides a quick start guide, including standard installation and an alternative method using `uv` for experienced Python users. It also offers optional bot interfaces for Telegram, QQ, Feishu, WeCom, and DingTalk, expanding its accessibility and integration capabilities. The project is licensed under the MIT License.
