Agentic API Adds a Stateful Agent Layer to vLLM

If you serve open models with vLLM, you already have fast token serving — but agent state, tool loops, and conversation history are still your problem. Agentic API is a stateful layer that moves that orchestration from the client into the server.

That means a client can continue from a previous_response_id instead of rebuilding and resending the full conversation state, while the gateway restores the session and keeps the loop running.

What Happened

Agentic API sits in front of vLLM and gives raw model inference an agent runtime. It adds the conversation state, tool-call loops, continuation across turns, streaming, and persistence that vLLM does not handle on its own.

Tool execution is explicit and split into three ownership categories:

• Gateway-owned tools like web search and MCP tools run server-side.
• Client-owned tools like Codex shell or editor actions return to the client.
• Provider-owned tools pass through to the model backend.

Why It Matters

That separation gives open models served by vLLM a stateful, OpenAI-compatible agent interface without forcing every application to build its own persistence and tool orchestration layer. The result is less glue code between model and agent loop, and a simpler path to production agents.