agent-evolver

What This Skill Does

The agent-evolver skill is a sophisticated autonomous learning engine designed for OpenClaw agents. It serves as a continuous improvement layer, enabling agents to transform execution failures into actionable insights. By leveraging Large Language Models (LLMs) and vector-based knowledge retrieval, the skill can analyze error logs, identify patterns, and store successful resolution strategies in a local SQLite database and ChromaDB vector store. Instead of repeating mistakes, an agent equipped with this skill can search for historically similar problems, understand the context of past failures, and apply refined execution strategies, effectively enabling the agent to 'evolve' its behavior over time.

Installation

To integrate the agent-evolver into your workflow, run the following command in your terminal:

clawhub install openclaw/skills/skills/lilei0311/agent-evolver

Once installed, ensure your OPENAI_API_KEY is configured in your environment variables to allow the skill to process error analysis and perform semantic vectorization of your experiences. You can further customize your evolution logic by modifying config/evolver_config.yaml to point to specific LLM models or local database paths.

Use Cases

• Error Resolution: Automatically diagnosing why a code execution or data task failed and generating corrective patches.
• Performance Optimization: Analyzing historical task data to suggest more efficient ways to handle complex multi-step instructions.
• Autonomous Learning: Building a persistent knowledge base of "Experience Capsules" that helps the agent avoid repetitive logical errors in tasks like numerical calculation or file handling.
• Regression Prevention: Retrieving past solutions to recurring edge cases, ensuring that once a problem is solved, the agent retains that capability permanently.

Example Prompts

1. "Agent, my last Python script failed with a KeyError. Please run agent-evolver to analyze the logs and suggest a fix based on previous successful runs."
2. "I want to improve the accuracy of your mathematical calculations. Use the agent-evolver skill to check our history for similar errors and refine your strategy."
3. "Show me the statistics of my agent's evolution over the last 20 tasks to see how much our success rate has improved."

Tips & Limitations

• Continuous Learning: For best results, trigger the evolution cycle after every non-trivial task; the more data you provide, the smarter the retrieval becomes.
• Data Privacy: Since this skill analyzes error logs and potentially sensitive task data, ensure your local evolution.db and ChromaDB paths are stored in secure directories.
• LLM Dependency: The quality of the "evolution" depends heavily on the model specified in the configuration. We recommend using high-performance models (e.g., GPT-4o) for complex logical errors.
• Initialization: Always ensure the EvolutionManager is initialized with a unique agent-id to prevent cross-contamination of experience databases if you are running multiple independent agent instances.