vLLM moves agent state into its serving layer

vLLM opened agentic-api on March 23 to move agent session state into its own serving layer. Most agent frameworks rebuild conversation history and multi-turn state on the client for every call. If vLLM handles that itself, frameworks built on top of it can drop that work.

The repo has 226 stars, but the sharper number is 62 forks against 46 open issues. That ratio reads as contributors extending the code, not spectators bookmarking it.

It's also first-party vLLM work, built by the same team that runs the inference engine. vLLM already handles a large share of production LLM inference. A stateful layer from the same team has a more direct path into existing deployments than a third-party framework would.

The repo doesn't say whether this ships as a core vLLM feature or stays a separate install. It has 21 open pull requests, so the API is still moving, and anyone building against it now should expect changes before it settles.