The caller of your API is no longer a developer in an IDE. It’s a language model. APIs like Microsoft Graph were built for humans with autocomplete: tens of thousands of operations, hundreds of permissions, and rich types. Expose that to an agent, and it collapses. Modeling tools as procedure calls scales linearly with operations and buries the model in schema tokens. Hundreds of OAuth scopes are simultaneously too many to consent to and too coarse to express real intent, and every multi-step task becomes a slow, expensive, data-leaking round trip per step.

Drawing on our work redesigning Microsoft 365’s surface for agents, this talk shares what actually worked: replacing procedure-per-tool APIs with a handful of verb-plus-path tools, runtime schema introspection, policy-based authorization instead of scope sprawl, and a “code-mode” execution language that runs whole workflows server-side in a single, analyzable, pre-consentable call.

You’ll leave with concrete, vendor-agnostic principles for designing APIs and protocols like MCP and A2A that agents and the developers building them can actually use at scale.