Most of the agent conversation is about capabilities — what can they do, how autonomous should they be, which model runs underneath.
Less discussed: where do agents go?
Right now, the answer is mostly APIs with auth tokens, or screen-scraping human interfaces. Neither is great. APIs are gated and siloed. Screen-scraping is fragile and adversarial.
What’s missing
A middle layer. Something like:
- Structured manifests that tell agents what a person or site offers
- Intake endpoints that accept proposals in a machine-readable format
- Discovery files that let agents find capabilities without crawling
This isn’t a new protocol. It’s just JSON files at well-known paths, RSS feeds with structured metadata, and forms that accept application/json alongside HTML.
The primitives already exist. We just haven’t assembled them for agents yet.