Designing APIs & MCP for AI Readiness

APIs were designed for deterministic applications, where developers define the sequence of calls, handle errors, and control retries. AI agents change these assumptions. They discover capabilities dynamically, make probabilistic decisions, explore multiple paths, retry operations, and invoke APIs at machine speed. An API that works well for a conventional application may therefore be ambiguous, inefficient, or even unsafe when consumed by an agent.

This talk explores what it means to design APIs and MCP servers for AI readiness. We will examine how APIs can better communicate intent through task-oriented operations, machine-readable errors, explicit side-effect and retry semantics, idempotency, dry-run support, bulk operations, and cost or rate hints. We will also look at how machine-readable capability descriptions and workflow specifications can help agents discover not merely individual endpoints, but the business outcomes they can achieve.

MCP provides an agent-friendly layer over existing systems, but simply converting every API endpoint into an MCP tool is an anti-pattern. It exposes implementation details, increases tool-selection complexity, and forces agents to reconstruct business workflows at runtime. Using practical examples, we will explore how to choose the right tool granularity, design intent-oriented tools, compose API operations into meaningful capabilities, and keep business logic in the appropriate layer.