# How to Build Safe, Scalable APIs for AI Agents

This talk explains why agents and MCPs depend on strong APIs, not magic, and what platform teams must do to support safe, reliable agentic use at scale. Drawing on Vanguard’s experience building external API and MCP platforms, Ted Hessing and Mike Wottle argue that AI raises delivery speed and the number of consumers while sharply increasing the cost of bad APIs, especially when non-human consumers can retry, fan out, and amplify errors. Their playbook centers on three connected practices: machine-readable and current API contracts such as OpenAPI specs and GraphQL schemas, automated guardrails at both build time and runtime, and operational discipline through standard work, feedback loops, and continuous improvement. They describe Vanguard’s automated spec generator, Baymax, which drafts specs for human review, and stress that governance, observability, resilience, policy enforcement, and intentional human-in-the-loop decisions remain essential as autonomous systems spread mistakes faster and farther.