microsoft/agent-lightning

The absolute trainer to light up AI agents.

What it solves

Agent Lightning is a lightweight framework designed to train AI agents using Reinforcement Learning (RL) without requiring significant changes to the agent's existing code or harness. It solves the difficulty of integrating complex agentic workflows—including tools, context, and control flow—into a training loop where the model is updated based on performance.

How it works

The framework uses a three-component architecture to bridge the gap between the agent's environment and the trainer:

  • Trainer: Manages the RL process using verl and vLLM to build training samples and update the model policy.
  • API Gateway: Acts as a proxy for model requests, capturing the interactions and trajectories that become training data.
  • Rollout Controller: Launches and manages the agents, either locally or as Kubernetes Jobs, allowing them to interact with their real-world environments.

Who it’s for

Researchers and developers building autonomous agents who want to apply RL to improve their performance in specific domains like coding, search, and mathematical reasoning.

Highlights

  • Minimal Code Changes: Agents can be trained with their real harnesses via a proxy, requiring zero changes to the agent's logic.
  • Lightweight Design: The entire framework consists of approximately 3,500 lines of code.
  • Kubernetes Native: Supports running agents as Kubernetes Jobs for scalable rollout collection.
  • Proven Results: Demonstrated a 14.6 percentage point gain on SWE-bench Verified for a Qwen3.5-9B coding agent using only 6K training samples.

Related

  • Project
    rllm-org/rllm

    An open-source framework for training language agents with reinforcement learning, allowing the same agent code to be used across different backends, sandboxes, and benchmarks.

  • Project
    Darwin-Agent/HarnessX

    HarnessX is a composable agent harness foundry that decouples model selection from behavior logic, allowing developers to evolve agent performance through automated harness optimization and RL fine-tuning.

  • Project
    verl-project/uni-agent

    Uni‑Agent is an open‑source framework that unifies agent harnesses, sandbox execution, and RL training, enabling thousands of concurrent long‑horizon LLM agents with reproducible recipes and benchmarked results.

  • Project
    InternLM/InternBootcamp

    InternAgentHarness (InternBootcamp) is an open‑source framework for creating, running, and evaluating multi‑turn, tool‑augmented LLM agents. It lets researchers define synthetic tasks, expose verified external tools, manage multi‑round interactions, compute rewards, and log full trajectories for data generation or RL training.

  • Project
    scaleapi/scale-agentex

    A framework for building and deploying AI agents that scales from simple chatbots to fully autonomous systems, supporting both synchronous and asynchronous workflows.