Your MCP Server is Bad (and you should feel bad) - Jeremiah Lowin, Prefect

The podcast centers on agentic product design for MCP (Message Compute Protocol) servers, emphasizing the need to curate interfaces optimized for AI agents rather than merely adapting human-centric APIs. It highlights three key differences between humans and AIs: discovery, iteration, and context. The discussion stresses the importance of focusing on outcomes over operations, advocating for tools designed around specific agent stories. The podcast also advises flattening arguments to simplify inputs, carefully crafting instructions and error messages to guide agent behavior, and respecting token budgets to avoid overwhelming the agent's limited context window. It cautions against directly converting REST APIs into MCP servers, suggesting it only as a bootstrapping method. The Q&A portion expands on topics like asynchronous background tasks, elicitation, and managing tightly coupled arguments.