Talk
Improving MCP tool schemas to increase agent reliability
As agents increasingly stitch together tools from many MCP servers, schema quality has become a critical factor in whether tool calls succeed at all. Agents rely entirely on tool schemas—not server code—to choose tools, construct arguments, and interpret results, making clear, expressive schemas essential for reliability. In this talk, we build MCP servers with FastMCP and Pydantic, consume them with Pydantic AI agents, and instrument runs with Logfire to observe tool selection and parsing in detail. We then evaluate tool-calling reliability and output correctness across models at scale using the Pydantic AI evaluation framework. Well-designed input and output schemas can both improve agent reliability, while also making our code more robust and type-safe. Win-win!
About
Pamela Fox is a human that loves to learn, teach, and create. She is currently a Python Cloud Advocate at Microsoft, where she helps developers use Python with Azure, Foundry, VS Code, and GitHub. Previously, she was an early developer advocate at Google, a full-stack developer at Coursera, and the CS content creator at Khan Academy. She also taught CS at UC Berkeley and for multiple non-profits around the bay area.