GenericAgent — Self-Evolving Agent with Skill Trees

Introduction

GenericAgent is a self-evolving AI agent that starts from a minimal seed of around 3,300 lines and progressively builds a skill tree through task execution. It learns reusable skills from completed tasks and applies them to future work, reducing token consumption by up to 6x compared to stateless approaches.

What GenericAgent Does

• Executes complex multi-step tasks across code, files, and system operations
• Builds and maintains a persistent skill tree from completed task patterns
• Reuses learned skills to handle similar tasks with fewer tokens
• Supports browser automation, desktop control, and shell execution
• Works with multiple LLM providers including Claude, Gemini, and GPT

Architecture Overview

GenericAgent uses a skill-tree architecture where each completed task can be distilled into a reusable skill node. The agent maintains a memory system that maps task descriptions to skill sequences. When a new task arrives, it searches the skill tree for applicable patterns, composes them into an execution plan, and only falls back to raw LLM reasoning for truly novel subtasks.

Self-Hosting & Configuration

• Install via pip with minimal dependencies
• Configure your preferred LLM provider via environment variables
• Set the workspace directory for file operations
• Customize the initial skill seed to match your domain
• Export and import skill trees between agent instances

Key Features

• Self-evolving skill tree that improves with each completed task
• 6x token reduction through skill reuse and pattern caching
• Multi-modal control: browser, desktop, terminal, and file system
• Lightweight 3.3K-line seed that grows organically
• Cross-provider support with Claude, Gemini, and OpenAI

Comparison with Similar Tools

• AutoGPT — general-purpose agent; GenericAgent focuses on skill-tree evolution
• CrewAI — multi-agent orchestration; GenericAgent is a single self-improving agent
• OpenInterpreter — code execution focus; GenericAgent adds persistent skill learning
• Evolver — GEP-based evolution; GenericAgent uses task-driven skill trees

FAQ

Q: What is a skill tree? A: A persistent graph of learned task patterns. Each node represents a reusable skill the agent discovered while completing previous tasks.

Q: How does it reduce token usage? A: By recognizing when a new task matches a known skill pattern, the agent skips the reasoning steps and executes the cached skill sequence directly.

Q: Can I transfer skills between projects? A: Yes. Skill trees can be exported as JSON and imported into other GenericAgent instances.

Q: Which LLM works best? A: Claude and Gemini models with large context windows work best for complex skill chains.

Sources

• https://github.com/lsdefine/GenericAgent