OPEN LINK ↗
// 2h agoOPENSOURCE RELEASE
Agent Lightning v1.0 makes agent RL practical
Microsoft’s Agent Lightning v1.0 is an open-source framework for training LLM agents through their existing deployment harnesses, including tool use, context management, and multi-agent workflows. Its reproducible coding-agent pipeline improved Qwen3.5-9B on SWE-bench Verified from 41.8% to 56.4%.
// ANALYSIS
Agent Lightning’s key contribution is treating the harness as part of the training system, tackling the messy retokenization and fragmented-trajectory problems that make real-world agent RL difficult.
- –Supports arbitrary agent harnesses while keeping execution and training modular
- –Addresses credit assignment, loss normalization, sample merging, and backend scheduling
- –Reports a substantial 14.6-point SWE-bench Verified gain using only 6,000 training examples
- –The roughly 3,500-line implementation lowers the barrier to reproducing and extending harnessed agentic RL
- –Open workflows and training scripts could make agent post-training more accessible beyond large research labs
// TAGS
agent-lightningagenttrainingtool-usetraining-infraopen-sourcebenchmark
DISCOVERED
2h ago
2026-08-21
PUBLISHED
2h ago
2026-08-21
RELEVANCE
10/ 10
AUTHOR
Discover AI