What does software look like when agents, not humans, are its primary users?

Most software today is designed human-first: dashboards, forms, clicks, visual state. Agents consume none of that. They read APIs, schemas and plain text — and they do it at a scale and speed where interface conventions designed for humans become friction, or worse, ambiguity.

The interesting version of this question isn’t “add an API to the existing product”. It’s: if the agent is the primary user, what does the product become? Structured, self-describing, forgiving of partial context? Does the human interface become a window onto agent state rather than the primary surface?

Early guesses, to be tested: machine-first interfaces reward explicitness over elegance, error messages matter more than features, and “documentation” stops being a page and becomes the contract itself. This is less a prediction than a direction to experiment in.