Agent Lightning: Decoupled RL Training for Agents

Agent Lightning is an open-source framework that decouples your agent runtime from the training loop. The core idea: your agent and the trainer never touch each other. The agent runs however it already runs (LangChain, AutoGen, CrewAI, OpenAI SDK, or plain Python) and emits spans as it goes. The trainer sits on the other side of a shared store and reads those spans. Neither side imports the other. This matters because turning an existing agent into something you can RL-train usually means rewriting it around the trainer's data model, or building the agent inside the trainer's framework from day one. Here the runtime stays put and the training loop is bolted on from outside. Mechanics: → every prompt, tool call, and reward is captured as a structured span, either through an emit_xxx helper you drop in or a tracer that picks them up automatically. → spans flow into a central LightningStore that holds tasks, resources, and traces. Both the agent runner and the algorithm read and write here, and that is the only contract between them. → the algorithm slot is pluggable. Reinforcement learning, automatic prompt optimization, and supervised fine-tuning all fit the same interface, and it can be an algorithm you wrote yourself. → in a multi-agent system you can point the trainer at one specific agent and leave the rest alone. The others still run and still emit spans, but only the selected one gets updated weights or prompts back. That last one is what I would actually reach for. Most agent RL work assumes you have one model to train. Real systems have a planner, a few specialists, a critic, and you only want to move the weights on the one that is actually the bottleneck. Being able to isolate that without pulling the whole graph apart is the difference between a research demo and something that runs in prod. Almost every agent framework shipping in 2025 was runtime-first. This one is training-first and treats which framework you built your agent in as an implementation detail rather than a requirement.