Microsoft Agent Lightning trains AI agents

Microsoft is trying to turn agent training into infrastructure, not a one-off research stunt. If the abstraction holds outside happy-path demos, this could become the default layer for agent self-improvement.

• –The core bet is decoupling: keep agent code intact while training runs on traced interactions, spans, and a separate optimization loop.
• –That lowers the adoption bar for teams with existing agents, since rewrites are usually what kill experimentation.
• –The project is moving beyond the paper: recent releases added a dashboard, REST API, Azure OpenAI support, MongoDB-backed store, and integrations like Tinker and Claude Code.
• –The real constraint is not the wrapper but the training loop itself: reward quality, debugging, and compute cost will decide whether this is practical in production.
• –Strong GitHub traction suggests developers are actively looking for agent RL tooling, not just better orchestration frameworks.