Microsoft Releases Agent Lightning v1.0, Boosting SWE-bench to 56.4%

Microsoft's Agent Lightning v1.0 introduces the Harnessed Agentic RL paradigm to train agents directly in their environments, closing the train-deploy gap and lifting SWE-bench performance to 56.4% with just 6K examples.

2026-08-19 ~ 2026-08-20 · 3 related posts

• Microsoft Releases Agent Lightning v1.0 for Reproducible Agent RL — microsoft · 2026-08-19
• Microsoft's Agent Lightning boosts SWE-bench score to 56.4% with 6K samples — omarsar0 · 2026-08-19
• Microsoft proposes Harnessed Agentic RL to bridge training-deployment gap — SergioPaniego · 2026-08-20