For three decades we've designed APIs for one kind of consumer: code written by a human who read the documentation before integrating. REST verbs, resource granularity, OpenAPI schemas, status codes, OAuth scopes, all of it assumes a developer makes the binding decisions at integration time and the running client merely executes them.

Agentic AI breaks that assumption. The consumer is now a non-deterministic reasoner that discovers capabilities, selects among them, and composes calls at runtime, with no human reviewing the request before it's made. An LLM doesn't read your docs; it reads your tool description, infers your parameters, and recovers, or doesn't, from your error messages. This session examines, concretely, what breaks at every layer of conventional API design when the caller is an agent, and the patterns that hold up under it.

We'll work through: why fine-grained CRUD surfaces push agents into brittle multi-step orchestration, and where intent-shaped operations outperform them; how schemas and descriptions become the agent's reasoning context, not just validation rules; designing error responses an agent can act on, retry, reformulate, escalate, rather than HTTP codes it merely logs; idempotency and confirmation patterns for a caller that may repeat or hallucinate a request, especially ahead of irreversible side effects; the authorization problem of delegated action, scoped capabilities, token binding, and the confused-deputy risk when an agent acts on a user's behalf across services; and async / long-running operation design for agents via MCP, tool-calling conventions, and durable execution.

Examples are drawn from building and operating an agentic platform in production at enterprise scale, an MCP gateway, a tool/skills registry, durable agent runtimes, but the patterns generalize to any API that expects autonomous-agent consumers, whether through MCP, native tool-calling, or direct invocation. Attendees leave with a checklist for auditing an existing API for agent-readiness, and a set of design defaults for new ones.