Agent Lightning Documentation

Welcome to the Agent Lightning v1.0 documentation. Start with the installation and quick-start guides, then use the configuration guides and examples below to build and train your own agents.

Agent Lightning v1.0 is a completely redesigned and reimplemented version with the following key features:

🪶 ~3,500 lines of core Python: Simplicity is the first principle.
🧩 Training with real agent harnesses: Agents interact with the model through the Agent Lightning v1.0 proxy with zero changes while keeping tools, context, control flow, and environments in the loop.
☸️ Native Kubernetes support: Agents run directly as Kubernetes Jobs without relying on external sandbox services.
💻 A complete coding-agent training example: The released pipeline covers data cleaning, reward-hacking prevention, and training scripts.

For the legacy Agent Lightning releases earlier than v1.0, see the v0.x code branch and the v0.3.0 documentation.

Getting Started

Guide	Description
Installation	Set up the base environment and the tested `verl` GPU stack.
Quick Start	Run a local end-to-end rollout-driven training job.
Basics	Learn the core components, rollouts, events, and trajectories.

Configuration

Guide	Description
Trainer Configuration	Configure `verl` integration, rollout collection, and trace aggregation.
API Gateway Configuration	Configure the API Gateway and model proxy.
Controller Configuration	Configure local and Kubernetes rollout runners.
Asynchronous Training	Configure collocated asynchronous collection and pause/drain behavior.

Examples

Example	Description
Calc-X	Train a math reasoning agent with AutoGen and MCP calculator tools.
GSM8K	Train an agent on grade-school math reasoning tasks.
ScienceWorld	Train an agent on interactive science tasks in a text environment.
Search-R1	Train a multi-turn retrieval and reasoning agent.
LLM-in-Sandbox	Train a general agent with computer and code execution tools.
Coding Agent	Train a coding agent using repository tests as feedback.