Agent Lightning v1.0 trains agents through their existing harnesses

Microsoft Research’s framework places a model API proxy between an existing agent and its training system.

Microsoft Research building 99 in Redmond, Washington.
Archive photograph, 13 February 2008: Microsoft Research building 99 in Redmond. Research-organization context for Agent Lightning; not the software launch. Reframed to the edition's 3:2 landscape hero from the archived original (1280x1920); WebP q88. Robert Scoble · CC BY 2.0

Microsoft Research has introduced Agent Lightning v1.0, an open-source reinforcement-learning framework designed to train agents using their existing execution harnesses. Instead of rebuilding the agent loop inside a trainer, the harness sends model requests through an LLM proxy.

Its API Gateway records model calls, while a Rollout Controller manages agent execution separately from a Customized Trainer. Rollouts can run as local processes or Kubernetes jobs. VCT’s takeaway: separating these responsibilities can reduce differences between training and deployment, but moving an API endpoint does not remove the need to review tool permissions, training data and compute requirements.

Source

[Microsoft Research announcement](https://www.microsoft.com/en-us/research/blog/agent-lightning-v1-0-a-3500-line-lightweight-agentic-rl-framework-for-training-agents-with-real-harnesses/).

Microsoft Research