MCP is all you need — Samuel Colvin, Pydantic

Video Description

Everyone is talking about agents, and right after that, they’re talking about agent-to-agent communications. Not surprisingly, various nascent, competing protocols are popping up to handle it. But maybe all we need is MCP — the OG of GenAI communication protocols (it's from way back in 2024!). Last year, Jason Liu gave the second most watched AIE talk — “Pydantic is all you need”. This year, I (the creator of Pydantic) am continuing the tradition by arguing that MCP might be all we need for agent-to-agent communications. What I’ll cover: - Misusing Common Patterns: MCP was designed for desktop/IDE applications like Claude Code and Cursor. How can we adapt MCP for autonomous agents? - Many Common Problems: MCP is great, but what can go wrong? How can you work around it? Can the protocol be extended to solve these issues? - Monitoring Complex Phenomena: How does observability work (and not work) with MCP? - Multiple Competing Protocols: A quick run-through of other agent communication protocols like A2A and AGNTCY, and probably a few more by June 😴 - Massive Crustaceans Party: What might success look like if everything goes to plan? ---related links--- https://x.com/samuel_colvin https://www.linkedin.com/in/samuel-colvin/ https://github.com/samuelcolvin https://pydantic.dev/ Timestamps 00:00:00 - Introduction: Speaker Samuel Colvin introduces himself as the creator of Pydantic. 00:00:42 - Pydantic Ecosystem: Introduction to Pydantic the company, the Pydantic AI agent framework, and the Logfire observability platform. 00:01:18 - Talk Thesis: Explaining the title "MCP is all you need" and the main argument that MCP simplifies agent communication. 00:02:05 - MCP's Focus: Clarifying that the talk focuses on MCP for autonomous agents and custom code, not its original desktop automation use case. 00:02:48 - Tool Calling Primitive: Highlighting that "tool calling" is the most relevant MCP primitive for this context. 00:03:10 - MCP vs. OpenAPI: Listing the advantages MCP has over a simple OpenAPI specification for tool calls. 00:03:21 - Feature 1: Dynamic Tools: Tools can appear and disappear based on server state. 00:03:26 - Feature 2: Streaming Logs: The ability to return log data to the user while a tool is still executing. 00:03:33 - Feature 3: Sampling: A mechanism for a tool (server) to request an LLM call back through the agent (client). 00:04:01 - MCP Architecture Diagram: Visualizing the basic agent-to-tool communication flow. 00:04:43 - Complex Architecture: Discussing scenarios where tools are themselves agents that need LLM access. 00:05:24 - Explaining Sampling: Detailing how sampling solves the problem of every agent needing its own LLM by allowing tools to "piggyback" on the client's LLM access. 00:06:42 - Pydantic AI's Role in Sampling: How the Pydantic AI library supports sampling on both the client and server side. 00:07:10 - Demo Start: Beginning the demonstration of a research agent that uses an MCP tool to query BigQuery. 00:08:23 - Code Walkthrough: Validation: Showing how Pydantic is used for output validation and automatic retries (model_retry). 00:09:00 - Code Walkthrough: Context Logging: Demonstrating the use of mcp_context.log to send progress updates back to the client. 00:10:51 - MCP Server Setup: Showing the code for setting up an MCP server using fast_mcp. 00:11:54 - Design Pattern: Inference Inside the Tool: Explaining the benefit of having the tool perform its own LLM inference to reduce the context burden on the main agent. 00:12:27 - Main Application Code: Reviewing the client-side code that defines the agent and registers the MCP tool. 00:13:16 - Observability with Logfire: Switching to the Logfire UI to trace the execution of the agent's query. 00:14:09 - Observing Sampling in Action: Pointing out the specific span in the trace that shows the tool making an LLM call back through the client via sampling. 00:14:48 - Inspecting the SQL Query: Showing how the observability tool can be used to see the exact SQL query that was generated by the internal agent. 00:15:15 - Conclusion: Final summary of the talk's points.