Microsoft released Agent Lightning v1.0 code, a Kubernetes rollout control plane and a complete coding-agent RL recipe.
Microsoft rebuilds Agent Lightning around real agent harnesses
2026-10-08Verified announcementTRIAL
Microsoft Research open-sourced a rebuilt Agent Lightning v1.0. An OpenAI-compatible proxy observes model calls from an existing agent harness, avoiding a rewrite of its execution loop inside the training framework. The roughly 3,500-line control plane runs local or Kubernetes rollouts; the authors report that about 6,000 samples raised Qwen3.5-9B from 41.8% to 56.4% Pass@1 on SWE-bench Verified.
What this means for you
For teams training real coding agents, the deployed harness can participate directly in RL, reducing divergence between training and production behavior. Results remain author-reported, and the system records prompts, responses and log probabilities; validate isolation, redaction, recovery and cost before adoption.
Evidence & related material
Technical details & scores
| Codebase | about 3,500 lines; open source |
|---|---|
| Integration | OpenAI-compatible LLM proxy observes an unchanged agent harness |
| Control Plane | API Gateway, Rollout Controller and customized verl trainer |
| Execution | local processes or native Kubernetes jobs; no paid sandbox dependency required |
| Scheduling | collocated asynchronous RL pauses new requests for model updates; author reports about 2x end-to-end speedup over synchronous RL |
| Coding Result | Qwen3.5-9B, about 6,000 SWE-smith samples, SWE-bench Verified Pass@1 41.8% to 56.4% |
| License | MIT repository |
Technical impact 4
Engineering value 5
Adoption signal 3
Maturity 3
Verification cost 4
Risk 3