Agent-callable API

Also known as: agent-native API, agent-first API, tool-callable API

An agent-callable API is one designed to be invoked by an autonomous LLM agent rather than a human developer. It exposes typed tool schemas, returns structured outputs an agent can parse, supports idempotency keys for safe retries, and handles long-running jobs through streaming or polling instead of blocking.

Example

Lamina's REST API ships an OpenAPI 3.1 spec, returns Zod-typed JSON, accepts an Idempotency-Key header on every write, supports DELETE on any job, and streams server-sent events on every long-running run.

Five properties of an agent-callable API

1) Typed tool schemas — OpenAPI 3.1 or JSON-Schema, ideally surfaced via MCP. 2) Structured outputs — every response is parseable JSON with a fixed shape, never free-form prose. 3) Idempotency keys — retrying a timed-out write returns the original result instead of double-spending. 4) Cancellable long jobs — the agent can kill any job and refund the credit. 5) Streamable progress — server-sent events let the agent report status back to the user without polling.

Why human-callable APIs fail under agents

Human-callable APIs assume someone will inspect the response, retry on a hunch, and pay attention to which call succeeded. Agents do none of that — they retry blindly, parse rigidly, and fail in unrecoverable ways when an API returns a 200 with a free-form text body. Agent-callable APIs treat retries, types, and structure as load-bearing.

References

Related terms