微软开源 Agent Lightning v1.0 智能体 RL 框架
Start here
微软研究院亚洲团队开源 Agent Lightning v1.0，提出 Harnessed Agentic RL 训练范式，让部署时使用的同一套 Agent harness 直接参与强化学习，无需在训练框架内重新实现 Agent。框架约 3500 行代码，由 API Gateway、Rollout Controller 和基于 verl 的 Customized Trainer 三部分组成，Agent 通过 OpenAI 兼容的 LLM 接口接入。
Generated by AI from reports · Updated 4 days ago
Progress timeline
- Microsoft ResearchSelectedMicrosoft Research open-sources Agent Lightning v1.0: a lightweight Agent RL framework with 3,500 lines of code.
Microsoft Research Asia team open-sourced Agent Lightning v1.0, proposing the Harnessed Agentic RL training paradigm, which lets the same agent harness used at deployment participate directly in reinforcement learning, without reimplementing the agent within the training framework. The framework is about 3,500 lines of code, consisting of three parts: API Gateway, Rollout Controller, and a Customized Trainer based on verl. Agents connect through an OpenAI-compatible LLM proxy, and the harness code remains unchanged. Agents run as standard Kubernetes jobs, can reuse self-managed clusters, cloud Kubernetes, or local infrastructure, and do not rely on paid commercial sandbox services. The team also proposed Collocated Async RL, which lets rollout and model updates share the same set of GPUs, achieving about 2x end-to-end speedup compared to synchronous RL in experiments.
Related discussion
Attention trend
New sources per day
- Oct 8New1 sources (1 outlets,0 accounts)
- Oct 9New0 sources (0 outlets,0 accounts)
- Oct 10New0 sources (0 outlets,0 accounts)
- Oct 11New0 sources (0 outlets,0 accounts)
- Oct 12New0 sources (0 outlets,0 accounts)
- Oct 13New0 sources (0 outlets,0 accounts)