AI Agents
NeMo-Agent-Toolkit
NVIDIA's toolkit for connecting, profiling and optimizing teams of AI agents — across any framework
★ 2.7k starsApache-2.0
View guide →Microsoft's trainer for AI agents - reinforcement learning for almost any agent framework
agent-lightning is an open-source ai agents project by microsoft (microsoft/agent-lightning on GitHub). It has 18,595 GitHub stars and 1,661 forks, is released under the MIT license, and was last updated on GitHub on 29 September 2026.
Choose an existing agent workload - the repo includes a full coding-agent training example.
Training setups are run from a source checkout.
Use uv sync and the provided setup script to prepare the training stack for your CUDA version.
Start from an included example, collect trajectories, and launch a training run.
Designed to train agents built with popular frameworks without rewriting them around the trainer.
Applies RL (with veRL integration scripts) to multi-step agent trajectories.
Ships an end-to-end example that trains a coding agent on SWE-bench-style tasks with data cleaning and reward-hacking prevention.
Keeps the training architecture to three lightweight components per the README.
The README reports measured gains for its examples (for example SWE-bench Verified improvements in the coding-agent pipeline).
Includes additional examples and a catalog of community projects built on it.
Get the source code from GitHub and enter the project directory.
git clone https://github.com/microsoft/agent-lightning.git
cd agent-lightningThe README's example installation targets a CUDA machine and uses uv plus the veRL setup script.
uv sync
bash scripts/setup_verl.sh 0.8.0 cu130 # example versions - see README installationOpen the examples in the repo (including the coding-agent training example) and adapt one to your agent.
# Browse the Examples section linked from the README and run one pipelineIt is maintained by Microsoft (Microsoft Research) as an open-source project on GitHub.
Yes - Agent Lightning is MIT licensed and free to use.
Training needs models to train and roll out: you bring your own model endpoints or local weights and the GPU capacity; the framework itself sells no API.
Compared with fine-tuning libraries like TRL, Agent Lightning is specialized for agents - trajectory-based training designed to wrap existing agent frameworks rather than single-turn LLM training.
Clone the repo, run uv sync, follow the installation guide for your CUDA setup, and run one of the included examples.
NVIDIA's toolkit for connecting, profiling and optimizing teams of AI agents — across any framework
Microsoft's GUI agent that grew into UFO³ — orchestrating agents across Windows, Linux and mobile devices.
The agent that grows with you — a Python framework for adaptive AI agents.
A Claude Code-style coding agent that runs anywhere and works with almost any model provider
Microsoft's R&D-Agent: LLM agents that automate data-driven research, Kaggle modelling and LLM fine-tuning.
Build and run agents you can see, understand and trust — framework plus a visual agent service
The most-starred other AI Agents projects in the LibreAI catalogue — the realistic shortlist if agent-lightning is not the right fit.
A personal AI that really does things — on any OS, with your data staying yours.
An agentic skills framework & software development methodology that works.
Agent harness optimization system with skills, memory and security for Claude Code, Codex and more
The agent that grows with you — a Python framework for adaptive AI agents.
DeepSeek's plugin-based agent harness where everything is a plugin.
Free and open — grab the source, star the repo, and start creating.