Agent Lightning v1.0: RL Data From Any Agent, No GPU

Microsoft shipped Agent Lightning v1.0 on August 17, 2026 — a full rewrite of its agent-RL framework, down to roughly 3,500 lines of core Python, MIT licensed, and already past 17k GitHub stars. The pitch that made it spread: train any agent with reinforcement learning, with almost zero code changes.

Nearly every writeup about it stops at the same place: "you need an A100 and the verl stack." That is true for the policy-update half. It is not true for the half that actually matters when you are starting out. The piece that turns your agent into something trainable — the API Gateway — is a plain FastAPI service that runs on a laptop CPU with no GPU, no vLLM, and no model weights.