Agents do not replace APIs; they consume them. Every agentic workflow bottoms out in an HTTP request to an endpoint that returns a response, with authentication, rate limits and error handling exactly as before. When people say agents change everything, what has actually changed is who is holding the client.
Two things are genuinely different, though, and both are worth taking seriously.
The consumer no longer reads. Everything the industry built to bridge the gap between an ambiguous contract and a working integration — tutorials, support forums, blog posts, a colleague who has done it before — does not exist for a machine. The contract has to carry the meaning on its own, which is a real and higher bar.
And the web was not built for this. It was built for humans, and a great deal of the connective tissue assumes one. That is why the ideas that never quite landed for human developers — semantics, hypermedia, self-describing affordances, machine-readable discovery, workflows — are suddenly relevant. Those were always the right answers to questions most people were not asking.
Also worth saying plainly: you often do not need an agent. Requests, cron jobs, webhooks and events handle an enormous amount of automation with less cost and far more predictability. Reach for the agent when the task genuinely requires judgment, not because it is the current shape of the conversation.
