# Agent Lightning — Reinforcement Training for AI Agents

> Open-source framework by Microsoft for training and evaluating AI agents with reinforcement learning, enabling self-improving agentic systems at scale.

## Install

Save as a script file and run:

# Agent Lightning — Reinforcement Training for AI Agents

## Quick Use
```bash
pip install agent-lightning
agent-lightning train --config examples/basic.yaml
```

## Introduction
Agent Lightning is an open-source framework from Microsoft designed to train AI agents using reinforcement learning. It provides a structured pipeline for reward modeling, policy optimization, and evaluation so teams can build agents that improve autonomously through interaction feedback.

## What Agent Lightning Does
- Trains agentic LLMs with RLHF and DPO-style reward signals
- Provides environment abstractions for multi-step task execution
- Supports distributed training across GPU clusters
- Integrates with popular model backends (Hugging Face, vLLM)
- Offers evaluation harnesses for measuring agent capability over time

## Architecture Overview
Agent Lightning follows a modular trainer-environment-evaluator architecture. The trainer orchestrates policy updates using configurable reward models, while environments expose step-based interfaces for tool use, code execution, or API interaction. Checkpoints and metrics flow through a central experiment tracker compatible with MLflow and Weights & Biases.

## Self-Hosting & Configuration
- Install via pip or clone the repository for development
- Define training configs in YAML (model, environment, reward)
- Requires CUDA-compatible GPUs for training workloads
- Supports multi-node setups via PyTorch distributed or Ray
- Environment variables control logging, checkpointing, and WandB integration

## Key Features
- Modular reward model architecture supporting custom scoring
- Built-in environments for code generation, web browsing, and tool use
- Scales from single-GPU experimentation to multi-node clusters
- Compatible with LoRA and QLoRA for efficient fine-tuning
- Tracks training runs with structured metrics and replay buffers

## Comparison with Similar Tools
- **TRL (Hugging Face)** — focuses on single-turn RLHF; Agent Lightning targets multi-step agentic loops
- **OpenRLHF** — strong on raw RLHF but lacks environment abstractions
- **Axolotl** — supervised fine-tuning oriented; no RL training loop
- **DeepSpeed-Chat** — lower-level; requires more manual orchestration

## FAQ
**Q: Does Agent Lightning require a custom reward model?**
A: No. It ships with default reward heuristics and supports plugging in external reward APIs or learned reward models.

**Q: Can I train on a single GPU?**
A: Yes, with smaller models and LoRA. Multi-GPU is recommended for full fine-tuning of 7B+ parameter models.

**Q: Which base models are supported?**
A: Any Hugging Face-compatible causal LM, including Llama, Mistral, Qwen, and DeepSeek families.

**Q: Is it production-ready?**
A: The framework is under active development. Microsoft uses it internally for agent research and releases updates regularly.

## Sources
- https://github.com/microsoft/agent-lightning
- https://github.com/microsoft/agent-lightning/blob/main/README.md

---
Source: https://tokrepo.com/en/workflows/asset-79d160b7
Author: Script Depot