Microsoft open-sources Agent Lightning v1.0: 3,500-line RL framework boosting SWE-bench +14.6 pts
Microsoft Research · rss · 2026-10-08
Microsoft Research Asia open-sourced Agent Lightning v1.0, built around a Harnessed Agentic RL paradigm where the deployment agent harness participates directly in RL training—no reimplementation needed.
Key points:
- 3,500 lines of code: API Gateway (OpenAI-compatible LLM proxy logging rollouts), Rollout Controller (Kubernetes jobs), and a verl-based trainer
- Simple adoption: point your harness's model endpoint at the proxy; harness code stays untouched
- Native Kubernetes support, no paid commercial sandbox dependency
- Collocated Async RL lets rollout and training share the same GPUs
- Solves retokenization/sample merging, rollout-level advantage calculation, loss normalization, and backend scheduling challenges
- Results: training Qwen3.5-9B on 6,000 open-source samples raised SWE-bench Verified Pass@1 from 41.8% to 56.4%, a 14.6-point gain
More from coding & agent
- Same model, very different performance: Dylan Patel on why the harness matters — MatthewBerman · 2026-10-08
- EreNodes after a year and 44 versions: prompts as clickable pills in ComfyUI — Erehr · 2026-10-08
- localmcplib: an open-source shared component library for local MCP servers — mostly_deterministic · 2026-10-08
- e2e: open-source agent turns natural-language goals into E2E tests for web and mobile — abhishek__AI · 2026-10-08
- TERMy24k: billed as the first local neuro-symbolic AI terminal assistant — gioscarab · 2026-10-08
- Managed Agents on Interactions API: one call spins up Antigravity agents in a remote sandbox — clmt · 2026-10-08