Microsoft opens Agent Lightning v1.0 for harnessed agentic RL

What did Microsoft Research publish?

arXiv 2608.17528, submitted 18 August 2026, presents Agent Lightning v1.0 as a lightweight framework for harnessed agentic RL in about 3,500 lines of code. Authors include Zhiyuan He, Siwei Zhang, Zhiwen Zhou, and collaborators listed on the Microsoft Research publication page.

The abstract says modern agents run inside harnesses that manage tools, context, and control flow, and that Agent Lightning connects arbitrary agents to RL through an LLM endpoint proxy—an approach the authors say later frameworks also adopted.

For a nearby open agentic-RL report, see Xiaomi’s MiMo-V2.6 RL write-up. For the benchmark itself, keep SWE-bench explained handy.

The Microsoft Research publication page lists the same author set and points to arXiv for the full text. Use the arXiv submit date (18 August 2026) when you need a day-level stamp; the MSR page itself shows only the month.

What is harnessed agentic RL?

The paper contrasts traditional agentic RL, where the training engine owns the environment loop, with harnessed agentic RL, where the harness owns that loop and the trainer observes sequences of LLM request-response pairs.

Authors list challenges that follow: retokenization, sample merging, advantage calculation, loss normalization, and backend scheduling. They present Agent Lightning v1.0 as a testbed for those issues rather than a claim that every challenge is solved for all harnesses.

Documentation for v1.0 highlights a trainer (verl and vLLM), an API gateway/proxy that captures trajectories, and a rollout controller that can run agents locally or as Kubernetes Jobs.

In practical terms, if your agent today talks to a model through an OpenAI-compatible base URL, the Lightning proxy aims to sit on that URL, log trajectories, and feed verl without rewriting the agent’s tool loop. Whether your harness’s streaming, tool-calling, and cancellation semantics survive that proxy is an integration question the docs and release notes address case by case.

What numbers does Microsoft report, and what is open?

On coding agents, the abstract reports that with 6K training examples and modest compute, RL improves Qwen3.5-9B on SWE-bench Verified from 41.8% to 56.4%, a 14.6-point absolute gain. That is the authors’ measured claim on their pipeline.

GitHub release v1.0.2 (29 September 2026) highlights MoE coding-agent training for Qwen3.5-35B-A3B with CISPO and R3 expert-routing replay, stating SWE-bench Verified rose from 47.8% to 61.6% after 1,792 training examples in the published pure-RL experiment. Again, that is the release note’s figure.

RL post-training vocabulary overlaps with classic preference tuning; for the older InstructGPT lineage see RLHF explained through InstructGPT research.

The repository is MIT-licensed. Docs describe CUDA 12.9/13.0 installs via uv and a setup_verl.sh helper. We did not run training.

v1.0.2 also records multimodal training fixes so image inputs from rollout traces reach VERL batches, plus a multimodal QA example. That is a separate capability track from the SWE-bench coding numbers; do not blend the two when citing results.

What should teams treat as limits?

The paper’s August date and the September v1.0.2 notes are different objects: use the paper for the paradigm and 9B coding pipeline; use v1.0.2 for the later MoE example and multimodal training fixes.

Reported SWE-bench numbers depend on harness, data cleaning, reward-hacking prevention, and compute the authors used. They are not transferable scores for your proprietary agent without re-running.

Kubernetes Job support and verl pinning are operational constraints. Windows native local runners are called out as failing fast in v1.0.2 notes for some paths—read the release before assuming a laptop demo.

Agent Lightning is a training stack, not a hosted product SLA. Availability of example scripts does not mean Microsoft hosts your rollouts.

The abstract’s claim that other frameworks later adopted proxy-based harnessed RL is historical context from the authors, not an endorsement ranking. Cite those frameworks from their own docs if you need comparative architecture detail.

What should a reader try first?

If you already run verl-based RL, start with the documented quick-start and a small non-SWE example before attempting the coding-agent pipeline.

Keep the deploy-time harness identical between training and serving if the whole point is harnessed RL; changing tools mid-flight defeats the proxy design.

Log token accounting and reward definitions explicitly. We are not auditing the authors’ anti-reward-hacking steps.

Document the exact proxy base URL, model name rewriting, and timeout behavior you use in production before you collect training traces. Mismatched retokenization between gateway and trainer is one of the failure modes the paper flags.

What did we not test?

We did not install Agent Lightning, run verl, launch Kubernetes jobs, or reproduce SWE-bench Verified scores. This article reports the arXiv abstract/page, Microsoft Research publication page, docs home, and GitHub release notes we opened.