maybe i am overthinking this but the more i look at AI agents using the web, the more the current web starts to feel kind of awkward.
like websites are still built assuming a human is sitting there, reading the page, ignoring the cookie popup, guessing which button matters, understanding which part is marketing and which part is actually useful.
but an agent does not really do that naturally.
it has to parse the page, figure out what is clickable, understand the form state, avoid random modals, compare options, maybe call tools, maybe retry when something fails, then somehow verify it actually did the right thing.
that sounds less like “browsing” and more like forcing software to cosplay as a human user.
which is probably fine for demos but idk how well that scales.
this is why all these things that seem separate to me are starting to feel connected. MCP, A2A, WebMCP, AI search, browser agents, bot traffic, agent security, all of it.
not saying they are the same thing obviously.
but they all point to the same pressure: software is becoming a real user of the web, not just humans.
and if that keeps happening then maybe websites need something beyond normal UI. not just better HTML or better accessibility, but some kind of agent-readable/action-readable layer.
basically not “AI kills websites” or anything dramatic like that.
more like websites keep existing for humans, but also need to expose themselves properly to machines.
kind of like SEO but instead of optimizing for search crawlers reading your content, you optimize for agents actually doing stuff.
not sure if this is a real architecture shift or just people putting new names on APIs again.
wrote a longer version in the attached medium post if anyone wants to read it.
[link] [comments]