Why LangChain Built Its MCP Support on FastMCP 4.0 | Sydney Runkle Interview

LangChain just shipped native MCP support, built on FastMCP v4. Sydney Runkle (Product Manager at LangChain, previously a maintainer at Pydantic) walks through why they built it that way. Jeremiah Lowin and Nate Nowack talk with Sydney about maintaining open source now that most pull requests are AI-generated, what actually changed in the MCP v2 spec, and the tool-name collision bug LangChain caught the night before a release. They also cover how to evaluate whether an MCP server is actually helping an agent, and close on Jev, the new model for fast, structured decisions. What we cover: - Maintaining open source repos in the age of AI-generated pull requests - What LangChain, LangGraph, and deep agents actually are - Why LangChain built its new MCP support on FastMCP v4 - What changed in the MCP v2 spec - Tool name collisions caught the night before a release - What's coming in the December MCP release - How to design evals for MCP servers - Jev, obviously Timestamps: 0:00 - What agent development looks like a year from now 0:21 - Welcome, and introducing Sydney Runkle 1:47 - Maintaining open source in the age of AI slop 4:41 - Borrowing Aztec's co-authorship trick for AI-era PRs 6:47 - Vercel's feedback issue type, and fixing anachronistic templates 8:14 - What LangChain, LangGraph, and deep agents actually are 10:47 - Does "deep" in deep agents mean persistent? 12:40 - Personal AI agents: Muse, and where trust breaks down 17:44 - Why LangChain built native MCP support on FastMCP v4 18:10 - What actually changed in the MCP v2 spec 22:04 - Fixing the client group abstraction for old and new servers 23:04 - Auth, caching, and session management wins 25:03 - The 11th-hour bug: tool name collisions across servers 27:48 - What's coming in the December MCP release 31:32 - How to design evals for MCP servers 33:17 - Two ways to evaluate whether an MCP server is any good 34:49 - The ID-formatting bug nobody can explain 35:44 - Building an evals framework from support tickets 37:17 - Jev: Fastino's new model for fast, structured decisions 39:20 - Using Jev for model routing and eval-as-a-judge 41:24 - Jev as a semantic router for tool selection 42:40 - Closing thoughts Resources: FastMCP on GitHub: https://github.com/PrefectHQ/fastmcp FastMCP docs: https://gofastmcp.com LangChain: https://www.langchain.com Try Prefect free: Prefect.io FastMCP is how teams build MCP servers. Prefect is how enterprises run them in production. If this was useful, subscribe. We love to talk about MCP, agent infrastructure, and production workflows in Python.