Agent Lightning
Microsoft's lightweight agentic RL framework that trains models through real agent harnesses unchanged, using a proxy gateway, verl/vLLM trainer and Kubernetes rollouts.
Tags
Description
v1.0 (open-sourced August 2026, ~3,500 lines) has three parts: a Trainer (verl + vLLM), an API Gateway that proxies model requests and captures training data, and a Rollout Controller that runs agents locally or as Kubernetes Jobs, so agents keep their own tools, context and environments. The repo reports a Qwen3.5-9B coding-agent run going from 41.8% to 56.4% on SWE-bench Verified with 6K samples, and Qwen3.5-35B-A3B from 47.8% to 61.6% with 1.8K samples (project-reported). Also ships an Agent Lightning skill for coding agents. Needs a CUDA GPU stack. Latest release v1.0.2 (2026-09-29).
Related
- SWE-bench: Benchmark and leaderboards (Verified, Lite, Multimodal, Multilingual) testing whether LLM agents can resolve real-world GitHub issues.
- SWE-Gym: Environment and dataset for training software-engineering agents and verifiers on real GitHub tasks (ICML 2025).
- SWE-smith: Toolkit and datasets for scaling training data for software-engineering agents by synthesizing task instances from GitHub repos (NeurIPS 2025 D&B).
Sources
- https://github.com/microsoft/agent-lightning
- https://arxiv.org/abs/2608.17528
- https://github.com/microsoft/agent-lightning/releases
Machine-readable
- JSON
- Markdown
- Source file on GitHub (edit via pull request)