An early experience with GPT-3 and JSON schema demonstrates how small API changes can drastically improve AI agent performance, cost, and consistency.
AI agents perform better and use fewer tokens when APIs are defined with detailed, machine-readable schemas like OpenAPI with clear constraints and examples.
Descriptive error messages with context, trace IDs, and suggested actions enable AI agents to self-correct and make more informed decisions.
Improve agent efficiency by chaining deterministic API calls into single tools and using progressive tool declaration to manage context overload.
Providing agents with an SDK or CLI offers a higher-level abstraction for interacting with APIs, similar to how human developers work.
Secure agent access by adopting industry standards like OAuth 2.1 with PKCE and understanding the risks of malicious MCP servers.
Common API failures include incomplete schemas, poor error handling, and a lack of discoverability, all of which hinder intelligent agent performance.
Jobs that call for the skills explored in this talk.