Originally published on AI Tech Connect.
What shipped Agent Lightning v1.0 landed on 17 August 2026, with a v1.0.1 patch release following shortly after. It is a complete redesign and reimplementation of the earlier Agent Lightning, at roughly 3,500 lines of core Python, with simplicity stated as the first design principle. The licence is MIT — the permissive end of the spectrum, and unusually clean for a training framework. Integration happens through a proxy. Your agent talks to the model through the Agent Lightning proxy with zero code changes, keeping its tools, context, control flow and environments in the loop. Agent execution and training are completely decoupled, and agents can run directly as Kubernetes Jobs without an external sandbox service. Microsoft reports that 6,000 training samples took an end-to-end Qwen3.5-9B…
Top comments (0)