What makes a service actually usable by AI agents?|

This is the most common failure point. Interactive login flows, CAPTCHAs, or non-standard OAuth implementations block autonomous access. AI agents need machine-friendly, documented authentication like OAuth 2.0 client credentials or static API keys. If a human is required in the loop, the agent cannot operate independently.

AgentMail is a clean example of getting this right. A single API call creates a new inbox. No browser redirect, no email verification, no human confirming anything. The agent authenticates with a key and proceeds.

AI agents rely on structured documentation to understand how to interact with an API. Providing a complete and well-structured OpenAPI specification, including descriptions, parameter details, and example responses, helps agents parse and reason about functionality without human interpretation.

This is more than just having docs. If your API isn't AI-ready, then it might as well not exist. An agent reads a schema and makes decisions based on what it finds. The richer and more precise that schema, the better an agent can reason about which endpoints to use and in what order.

Returning plain-text errors or non-standard error formats makes it difficult for agents to detect and recover from issues. AI agents benefit from consistent, structured errors with codes, types, and resolution hints. Vague messages force unnecessary guesswork or retries.

An agent that hits an error needs to decide what to do next: retry, fall back, escalate, or abort. That decision requires structured signal. {"error": "rate_limit_exceeded", "retry_after": 30} gives it everything it needs.

Agents don't poll. Or rather, they shouldn't have to. A service that requires an agent to repeatedly check for state changes is a service that creates unnecessary overhead and introduces race conditions. The best agent-ready services push events to the agent via webhooks, WebSockets, or server-sent events the moment something relevant happens.

Stripe has been ahead of this curve for years, and AgentMail delivers incoming email events in real time via webhooks and WebSockets for the same reason.

Anthropic's Model Context Protocol and Google's Agent-to-Agent Protocol are establishing the equivalent of HTTP-level standards for agentic AI, enabling any agent to use any tool or collaborate with any other agent. Services that expose an MCP server make themselves natively discoverable and callable by any agent that speaks the protocol, without custom integration work.

Services that have shipped MCP endpoints are, in effect, opting into the agentic economy's default discovery layer.

A newer but fast-growing convention: a plain-text file at /llms.txt on a service's domain that gives AI systems a concise, structured summary of what the service does, what its APIs look like, and how to get started. Think of it as a robots.txt for agents.