Reasoning models and tool-calling agents are now central to production LLM systems, but the infrastructure inside vLLM that enables them is rarely documented outside the codebase. This talk opens that black box.