Microsoft's Agent Lightning just got a proper 1.0. The project started as a research framework for training AI agents with reinforcement learning; for v1.0 it was refactored almost from scratch into a roughly 3,500-line framework built around one idea: you should be able to train an agent's underlying model without modifying the agent code itself.
The team calls the concept "harnessed agentic RL," and the technical report spelling it out was posted to arXiv on August 181. Here's what shipped and why it's worth a look.
The idea: the harness stays in the loop
Most agent training pipelines bolt the training loop around the agent. Agent Lightning flips that. Instead of making the training engine drive the environment interaction, the deploy-time harness keeps owning that loop, meaning tool calls, context, and control flow, while the trainer only ever sees the LLM's request-response pairs.
In practice that means the framework inserts an OpenAI-compatible proxy between the agent and the model it calls, captures the traffic, and converts it into training data. The agent keeps using its real harness unchanged, and the training side stays decoupled2.
The design itself is not new. The original Agent Lightning already introduced this disaggregated, proxy-based architecture, and the technical report notes that the approach has since been picked up by other frameworks, including verl Uni-Agent, AReaL 2.0, slime, and Polar1. The v1.0 rewrite is a consolidation of that approach into a much smaller codebase with the harness story made central.
What v1.0 looks like
The refactored architecture is three components2:
- Trainer runs
verland vLLM, builds training samples, and updates the policy. - API Gateway proxies model requests and captures the training data.
- Rollout Controller runs agents locally or as Kubernetes Jobs.
The Rollout Controller launches agents, and the Gateway turns those interactions into training material. Kubernetes support is native, so agents can run directly as Jobs rather than depending on an external sandbox service2.
Simplicity is the stated goal. The README describes it as a lightweight agentic RL framework at about 3,500 lines, treating simplicity as a first principle2. It's MIT-licensed, and the Python package is agentlightning23.
The headline numbers
The most practical claim is a coding result. Using only 6K training samples and modest compute, pure RL moved a Qwen3.5-9B workflow on SWE-bench Verified from 41.8% to 56.4%, a gain of 14.6 percentage points12. The full pipeline is released, including data cleaning, reward-hacking prevention, and training scripts, so the result is meant to be reproducible rather than just impressive.
The paper also evaluates the framework on instruction-following and search agents, not only coding. It ships example pipelines for math reasoning (Calc-X, GSM8K), a text-based science environment, multi-turn retrieval (Search-R1), and a general agent with computer and code execution tools2.
A separate, quieter release
Six days after the technical report went up, the maintainers cut v1.0.1 on August 24, and this one is a different flavor of "training." It marks the first official release of the Agent Lightning Skill, a Claude Code-style skill meant to help coding agents optimize other AI agents, guiding systematic changes to prompts, tools, workflows, models, and reasoning settings while balancing accuracy, cost, latency, and reliability through measured iteration4.
It installs into Claude Code, Codex, or GitHub Copilot with a single command4:
gh skill install microsoft/agent-lightning agent-lightning --agent <agent>gh skill install microsoft/agent-lightning agent-lightning --agent <agent>So there are effectively two things named "Agent Lightning" now: the RL training framework (v1.0) and the skill-based agent optimizer (v1.0.1). The technical report anchors the former; the release notes cover the latter.
Why this matters
This is worth watching because it's a concrete attempt to make RL training a practical step for people building agents rather than a research-only pursuit. The guarantee of zero changes to your agent is what makes it testable. If you can point it at an existing harness and get measurable improvement without rewiring anything, that's a much lower barrier than the frameworks that came before.
The project page has been live since June 202525, and the codebase carries 17.9k stars, so this is not a stealth launch. The new part is the v1.0 tag: a clean consolidation point, three simple components, and a reproducible coding-agent result attached to it.
If you build agents and have been putting off RL fine-tuning because the harness integration looked like a project in itself, this is worth a weekend to test.
- Technical report, Agent Lightning v1.0: Towards Harnessed Agentic RL, arXiv:2608.17528.
- Agent Lightning repository README.
- Agent Lightning, Microsoft Research project page.
- Agent Lightning v1.0.1 release notes.
- Microsoft Research Agent Lightning project page (resources dated June 2025).