Microsoft · 2026-08-17 · major
Agent Lightning v1.0 — Microsoft's RL trainer plugs into real agent harnesses
Agent Lightning v1.0 trains an agent with reinforcement learning while the agent keeps running inside its own harness. Microsoft rewrote the framework down to about 3,500 lines of core Python and added native Kubernetes support.
Agent Lightning v1.0 lets you reinforcement-learn a model without rewriting the agent that will run it in production.
Quick facts
| Maker | Microsoft |
|---|---|
| Version | v1.0.0 |
| License | MIT |
| Core code | ~3,500 lines of Python |
| Training stack | verl + vLLM |
| What's new | Full rewrite, harness-owned rollouts, Kubernetes Jobs |
| Paper | arXiv 2608.17528 |
Benchmarks
| Qwen3.5-9B after Agent Lightning RL | 56.4% | |
|---|---|---|
| Qwen3.5-9B before training | 41.8% |
What is it?
Agent Lightning v1.0 is a full rewrite of Microsoft's open-source reinforcement learning framework for AI agents, released on August 17, 2026 under the MIT license. The new version drops to about 3,500 lines of core Python and adds native Kubernetes support, so each agent rollout can run as a Kubernetes Job instead of needing a separate sandbox service.
How does it work?
The paper calls the approach "harnessed agentic RL": the deploy-time harness, not the training engine, owns the loop that talks to the environment. Your agent points at an Agent Lightning proxy instead of the model endpoint, and the proxy records the calls it needs for training while the agent keeps its own tools, context and control flow. The framework then handles retokenization, sample merging, advantage calculation, loss normalization and backend scheduling on a verl and vLLM training stack.
Why does it matter?
Most RL-for-agents work makes you rebuild the agent as a training environment, and the rebuilt version behaves differently from the one you ship. Training through the real harness removes that gap, and Microsoft's own run shows it pays: 6,000 training samples moved Qwen3.5-9B from 41.8% to 56.4% on SWE-bench Verified. Teams already running AutoGen or MCP-based agents can post-train a smaller open model instead of paying for a frontier one.
Who is it for?
ML engineers post-training agent models
Frequently asked questions
- Do I have to change my agent's code to train it with Agent Lightning?
- No. Agent Lightning v1.0 works by putting a proxy between your agent and the model endpoint, so the agent runs unmodified and keeps its existing tools, context handling and control flow. Microsoft describes this as training with real agent harnesses, which is the headline change in v1.0 over the earlier v0.x design.
- What changed between Agent Lightning v0.3 and v1.0?
- Agent Lightning v1.0 is a complete refactor rather than an incremental update. Microsoft rewrote the codebase to about 3,500 lines of core Python, moved rollout ownership to the deploy-time harness, added support for the latest models, and shipped native Kubernetes support so agents can run as Kubernetes Jobs without an external sandbox dependency.
- What infrastructure do I need to run it?
- Agent Lightning v1.0 trains on a verl and vLLM GPU stack, and the repository ships a setup script for verl. For scale-out, the release adds native Kubernetes support so each agent rollout becomes a Kubernetes Job. That means a GPU cluster is the realistic baseline; this is a post-training framework, not something you run on a laptop.
- How big was the SWE-bench gain, and on what?
- Microsoft reports that Agent Lightning v1.0 raised Qwen3.5-9B from 41.8% to 56.4% on SWE-bench Verified, a gain of 14.6 percentage points, using only 6,000 training examples. The paper also covers instruction-following and search agents, so the coding result is one of three agent types tested rather than the only one.
Try it
git clone https://github.com/microsoft/agent-lightning && uv sync