Understand the project
Practical overview
When it fits
Use it when the training process needs to observe and learn from an agent operating through its real harness, rather than only tuning standalone model responses. It is also relevant when rollouts need to run locally or as Kubernetes Jobs.
Who it is for
- Developers building reinforcement-learning workflows for AI agents.
- Teams training LLM-based agents with tools, environments or repository tests.
- MLOps practitioners integrating agent rollouts into local or Kubernetes-based infrastructure.
Documented uses
- Train an existing AI agent while retaining its tools, context, control flow and environment.
- Collect agent interactions through a model proxy for reinforcement-learning workflows.
- Run coding-agent, search, sandbox and interactive-environment training examples.
- Organize trainer, gateway and rollout-controller components in an agent-training system.
Before you choose
Requirements & limits
The repository is public, active and MIT-licensed. The documented installation example targets a CUDA 13.0 machine and uses uv with a verl 0.8.0 cu130 setup.
- The framework is Python-based and its documented trainer stack uses verl and vLLM.
- Its API Gateway preserves model-request interaction data for training.
- The Rollout Controller supports local execution and Kubernetes Jobs.
- The examples span math, science, retrieval, sandbox and coding-agent workloads.
- The documented installation path is GPU-oriented, with a CUDA 13.0 example.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Evidence and transparency
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 2 October 2026
- Original website title
- microsoft/agent-lightning · Official homepage title · retained in its source language