Microsoft Agent Lightning

Agent Lightning is Microsoft Research's open-source framework for improving production AI agents with reinforcement learning. It sits behind an API proxy so agents built on LangChain, AutoGen or the OpenAI Agents SDK run unmodified while their traces are captured and turned into training data, letting platform teams raise task accuracy on their own workloads instead of waiting for a better base model.

Agent Lightning is an open-source, MIT-licensed framework from Microsoft Research that applies reinforcement learning to AI agents without requiring those agents to be rewritten for training. Version 1.0 shipped on 17 August 2026 as a complete rewrite of the earlier v0.x line, reducing the core to roughly 3,500 lines of Python organised around three components: a Trainer that drives the verl and vLLM GPU stack and aggregates traces into training samples, an API Gateway that proxies model requests and captures the data, and a Rollout Controller that executes agents locally or as native Kubernetes Jobs. The central design claim is complete decoupling of agent execution from training: because the agent talks to the model through the proxy, it keeps its own tools, context handling and control flow, and integrating an existing agent built with LangChain, the OpenAI Agents SDK, AutoGen or from scratch requires almost no code modification. The underlying method, published as LightningRL in the accompanying arXiv paper by Xufang Luo and colleagues, formulates agent execution as a Markov decision process and adds a hierarchical credit-assignment module that decomposes complex multi-step trajectories — including multi-agent ones — into individually trainable transitions. The paper evaluates text-to-SQL, retrieval-augmented generation and math tool-use tasks, reporting stable continuous improvement across all three. The headline result cited for v1.0 is a 14.6-point absolute gain on SWE-bench Verified for Qwen3.5-9B, from 41.8% to 56.4%, using only 6,000 training examples and modest compute. The project has approximately 17.9k GitHub stars and 1.6k forks with around 99 open issues, and v1.0.1 introduced the Agent Lightning Skill, which lets coding agents optimise other agents.

Category

Agent Development

Pricing

Free

Target Market

CTOs, Heads of AI, ML Engineers, Platform Engineers, AI Researchers

Deployment

Open-source, Self-hosted

Founded

1975

Headquarters

Redmond, United States

Team Size

500+

• ✓
  Zero-code-change agent integration

  Agents communicate through an API proxy, so existing LangChain, OpenAI Agents SDK or AutoGen implementations are trained unmodified while retaining their tools and control flow.

• ✓
  LightningRL hierarchical credit assignment

  Formulates agent execution as a Markov decision process and decomposes long multi-step trajectories into individually trainable transitions, which is what makes multi-turn and multi-agent RL tractable.

• ✓
  Three-component architecture

  Trainer, API Gateway and Rollout Controller keep the system auditable at roughly 3,500 lines of core Python rather than a sprawling training platform.

• ✓
  Native Kubernetes rollouts

  Agent rollouts execute as Kubernetes Jobs directly, removing the need for a separate external sandbox service in the training loop.

• ✓
  verl and vLLM training stack

  Builds on established open GPU infrastructure for rollout generation and policy optimisation instead of a bespoke trainer, easing operational integration.

• ✓
  Agent Lightning Skill

  Introduced in v1.0.1, this lets a coding agent drive the optimisation of other agents, turning post-training itself into an agent-assisted workflow.

• ✓
  MIT licence

  Fully permissive for commercial use and modification, with no field-of-use or revenue restrictions to clear in procurement.

• •
  Cost-down of a frontier-model agent

  Post-train a 9B open model on captured production traces until it matches the incumbent API agent on the specific task, then switch.

• •
  Internal text-to-SQL accuracy lift

  Train a self-correcting SQL agent against the organisation's real schemas and query logs, raising first-attempt correctness on analyst questions.

• •
  Repository-specific coding agent

  Tune a coding agent on an internal monorepo so it follows house conventions and passes the existing test suite more reliably.

• •
  RAG relevance tuning

  Use organisational relevance judgements as reward signal to improve retrieval and synthesis behaviour on a proprietary knowledge base.

• •
  Multi-agent workflow optimisation

  Apply hierarchical credit assignment across a multi-agent pipeline to identify and improve the step that actually causes task failure.

• •
  Benchmark-driven agent research

  Evaluate RL post-training approaches against real agent harnesses with tools and environments intact rather than simplified proxies.

The framework itself is free under the MIT licence with no seat, usage or commercial restriction, so every dollar of cost is infrastructure rather than software. The real spend is GPU capacity for the verl and vLLM training stack plus rollout execution, and the project is explicit that a GPU cluster is the realistic baseline — laptop deployment is not practical. Microsoft publishes no paid tier, no hosted service and no support contract for it, so budget for cluster time, CUDA-compatible hardware and the engineering effort to build a reward signal, which is usually the larger cost.

This page was written from 5 sources, 4 on domains other than microsoft.github.io.

1. 1.github.com — agent lightning
2. 2.microsoft.github.io — latestvendor
3. 3.arxiv.org — 2508.03680
4. 4.ai-tldr.dev — microsoft agent lightning 1 0
5. 5.hn.algolia.com — hn.algolia.com