ODSC AI East 2026 Keynote with Prefect's Jeremiah Lowin: Your MCP Server is Bad

Model Context Protocol is positioned as a “USB-C for AI” that standardizes how agents connect with traditional software. Many implementations are weak because developers wrap existing REST APIs, which forces agents to repeatedly rediscover API structure and documentation. This repeated discovery creates high token usage, increased latency, and higher costs. Developers are urged to stop optimizing for human interfaces alone and instead design an “agentic UI” that reduces the agent’s decision space. Better MCP servers curate the surface area exposed to the agent. Tool search is presented as a way to avoid loading every tool definition at handshake, enabling the agent to find relevant tools without consuming large context windows.

"Lowin argues that the industry must move beyond "dumbing down" servers and instead embrace new architectural patterns that prioritize the specific needs of AI agents. Many developers currently create suboptimal Model Context Protocol (MCP) servers by simply wrapping existing REST APIs. While REST principles favor broad, atomic endpoints for human developers who can easily "stitch" workflows together, AI agents face unique structural challenges."

"Most notably, agents suffer from "amnesia," meaning they must rediscover the API's structure and documentation with every interaction. This discovery process is incredibly expensive when using standard REST wrappers. Every interaction requires the agent to consume significant tokens to understand the context, leading to high latency and increased costs. Lowin suggests that developers often obsess over user interfaces (UI) for humans but neglect the "agentic UI" required for AI to function efficiently."

"To build better Model Context Protocol (MCP) servers, developers must reduce the decision space and curate the surface area exposed to the agent. To address the limitations of traditional tool registries, Lowin highlights "tool search" as a critical innovation. In earlier implementations, MCP clients often stuffed every available tool definition into the agent's context window upon handshake. This practice frequently consumed tens of thousands of tokens, leaving little room for actual task processing."

Read at Medium