AI × AEC · CURATED INSIGHT
Microsoft Research Asia Open-Sources Agent Lightning v1.0
Microsoft Research Asia has introduced Agent Lightning v1.0, a 3,500-line reinforcement learning framework that trains agents using their existing deployment harnesses directly via an LLM proxy.
Agent Lightning v1.0 implements a Harnessed Agentic RL control plane across an API Gateway, Rollout Controller, and Customized Trainer. The open-source system runs agents natively as standard Kubernetes jobs without commercial sandboxes and introduces Collocated Async RL to share GPUs. In tests, a coding pipeline raised Qwen3.5-9B Pass@1 on SWE-bench Verified from 41.8% to 56.4% using about 6,000 samples.
Key takeaways
- The framework places an LLM proxy between agent harnesses and models, running agents as Kubernetes jobs on local or cloud clusters without paid sandbox services.
- Using collocated async RL and about 6,000 training samples, the pipeline lifted Qwen3.5-9B Pass@1 on SWE-bench Verified from 41.8% to 56.4%.
Source：Microsoft Research Read the original ↗
Prepared, checked and automatically published by AI; not individually reviewed by a human.