A Microsoft Research framework for optimizing AI agents through Reinforcement Learning and prompt tuning with zero code changes.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install agent-lightning
2. Manual Installation
Copy the skill folder to one of these locations
~/.openclaw/skills/ <project>/skills/ Priority: Workspace > Local > Bundled
3. Prompt Installation
Copy this prompt to OpenClaw to install it automatically.
Help me install agent-lightning using Clawhub. If Clawhub is not installed, install it first (npm i -g clawhub).
Prefer to download?
Get the raw skill files in a ZIP archive.
What is Agent Lightning?
Agent Lightning is a specialized framework developed by Microsoft Research that allows developers to turn standard AI agents into optimized powerhouses. By employing Reinforcement Learning (RL), Automatic Prompt Optimization (APO), and Supervised Fine-tuning (SFT), it provides a pathway to enhance agent logic and efficiency without rewriting core code. This framework is a critical addition to the Openclaw Skills ecosystem for developers who need to scale agent capabilities across various platforms like LangChain, CrewAI, and AutoGen.
The tool focuses on selective optimization, meaning you can choose to improve specific agents within a complex multi-agent system. Its zero-code-change philosophy, implemented via tracers or simple helper functions, ensures that your existing development workflow remains intact while the underlying intelligence of your agents evolves through automated training cycles.
Agent Lightning Use Cases
- Scaling agent performance in complex reasoning tasks such as math and coding using Reinforcement Learning.
- Refining system prompts automatically to eliminate hallucination and improve task adherence.
- Training specialized SQL agents to handle complex database schemas with higher precision.
- Optimizing multi-agent systems where specific agents require fine-tuned behavior for collaborative success.
- Integrating agent training workflows into the Openclaw Skills library for centralized management.
How Agent Lightning Works
- Instrument the agent by adding emit helpers or using the built-in tracer to capture interaction data without modifying the agent's core logic.
- Define a YAML configuration file to specify the training algorithm (such as GRPO or APO), training episodes, and evaluation benchmarks.
- Execute the training command to run the agent through the specified environment and collect performance traces.
- Apply optimization algorithms like Group Relative Policy Optimization to iteratively improve the agent's policy or prompts.
- Evaluate the agent against held-out tasks to ensure generalization and avoid overfitting.
- Export the optimized prompts or model weights for production deployment within your Openclaw Skills environment.
Agent Lightning Setup
To install the framework, use the following command in your terminal:
pip install agentlightning
For developers requiring the latest nightly builds to test with new Openclaw Skills features:
pip install --upgrade --index-url https://test.pypi.org/simple/ --extra-index-url https://pypi.org/simple/ --pre agentlightning
Ensure your environment variables are configured for the trainer:
export OPENAI_API_KEY="sk-..."
export AGL_STORAGE="s3://your-bucket/path/"
Agent Lightning Data Schema & Taxonomy
Agent Lightning utilizes a structured approach to manage training data and agent state, making it easy to integrate with other Openclaw Skills.
| Data Type | Description | Storage Method |
|---|---|---|
| Traces | Captured logs of messages and tool calls | LightningStore |
| Configs | YAML files defining hyperparameters and tasks | Local Filesystem |
| Checkpoints | Saved states of prompts or model weights | Local/S3 |
| Logs | Detailed training metrics and loss curves | ~/.agent-lightning/logs/ |
| Evaluations | Results from benchmark tasks like math or coding | JSON Reports |
Agent Lightning Advanced Features
- Universal compatibility with major agent frameworks including OpenAI Agent SDK and Microsoft Agent Framework.
- Support for Group Relative Policy Optimization (GRPO) for stable agent policy improvements.
- Automatic Prompt Optimization (APO) for programmatic refinement of system instructions.
- Integrated dashboard for real-time monitoring of training convergence and loss curves.
- LightningStore for synchronized management of tasks, resources, and traces across the Openclaw Skills landscape.
- Remote storage integration for managing large-scale training artifacts on S3.
SKILL.md
Loading
Related Openclaw Skills
Agent Guardrails
A mechanical enforcement framework that uses git hooks and scripts to ensure AI agents follow project standards and security protocols.
4090 Server Remote Control
A specialized skill for remote SSH management of high-performance 4090 servers hosting Dify and Docker-based AI infrastructure.
TikTok App Marketing Automation
A comprehensive AI agent skill designed to automate the entire TikTok slideshow marketing pipeline from competitor research to conversion-driven iteration.
Zen+ Health
A professional wellness integration for Openclaw Skills that provides mindfulness exercises, mood tracking, and burnout prevention tools.
Grove Tipping CLI
A CLI-driven micro-payment tool that enables AI agents and humans to send quality signals through financial tipping.
Psyche Engine
An emergent psychological affect system that transforms interaction data into complex internal states and behavioral snapshots for AI agents.
Featured*
