getting-started.md

Getting Started

Prerequisites

• Python 3.9+ (the kernel needs nothing else)
• Optional: fastapi + uvicorn for the control plane (pip install -r requirements.txt)

Installation

git clone https://github.com/huzjie/agentlightning.git
cd agentlightning
python -m venv .venv
source .venv/bin/activate   # Windows: .venv\Scripts\activate
pip install -r requirements.txt

First run

python -m agentlightning doctor

Expected output lists the harnesses (codex, deepseek, claude_code, local, mock), providers (9 + mock), and reward functions (5).

Train in mock mode

python -m agentlightning train --task "Solve: implement fizzbuzz"

This collects trajectories in the mock harness, scores them, and runs PPO over the policy's learnable parameters. Fully offline.

Rollout

python -m agentlightning rollout --task "Answer: what is 2+2?" --episodes 16

Verify

python -m agentlightning verify --task "capital of France" --n 5 --mode majority

Serve

python -m agentlightning serve --host 0.0.0.0 --port 8000