MCP made tool calling the default way agents interact with the world, but a growing pattern called "Code Mode" has agents write and execute code that calls MCP tools programmatically instead of invoking them one at a time.

This talk breaks down the real tradeoffs: when tool calling wins (simplicity, auditability, guardrails) and when Code Mode wins (composability, fewer round trips, handling large tool surfaces without blowing the context window). Using concrete examples, we'll look at where each pattern breaks down in practice and give developers a framework for deciding which to reach for as they design their own MCP-based agents.