Choosing an agent framework

You have now built Relay's support agent three times: as a loop you wrote by hand in module eight, with the OpenAI Agents SDK in module ten, and with LangChain and LangGraph in this module. The tools were the same three functions each time. What changed was who owned the loop, and how much of its shape you were allowed to draw.

That is the whole decision. An agent framework is a way of running the model-tool-model loop, plus some services around it: memory, streaming, tracing, multi-agent wiring. The frameworks differ in how much of the loop they let you see and change, and in which extras come in the box. MCP, the subject of the next module, is not a fourth framework; it is a way to make the tools independent of whichever one you pick.