Last week, the Safari team formally expressed their opposition to the WebMCP API. In Safari’s explanation of why they oppose WebMCP, they captured something I’ve been trying to articulate.

In the comment explaining the Safari team position, Mike Wyrzykowski writes:

When a site’s actions are hard for an agent to use, that is a gap in the page’s own semantics, and the fix, in our opinion, is to close it in the platform’s shared layers (HTML and ARIA), where the user, assistive technology, and agents all benefit. 

Amen.

In fact, this sentiment should guide any efforts to change the web to support AI agents better. It’s possible that not every enhancement for AI can be structured in a way that supports all of the web’s shared layers, but we should strive to do so first before reaching for solutions that only benefit AI and not the humans the web is supposed to serve.

The Safari teams continues:

Our deeper concern is architectural. An agent acting on a user’s behalf is, in effect, assistive technology: it should operate a site as the user would, and the site should not single it out for different treatment. 

And on accessibility:

The proposal says WebMCP “is not designed for ingestion by accessibility technology”, the same fork: richer, actionable semantics reach agents while screen-reader and keyboard users get less.

All of this reminds me of a recent discussion on the Web Incubator Community Group (WICG) mailing list about a proposed standard to help AI understand domain authority. The proposed standard is designed to solve the following problem:

AI assistants often misidentify or misrepresent domains because there is no consistent, machine-readable, domain-controlled source of identity data. Today, models rely on scraped pages, inconsistent metadata, third-party aggregators, or outdated indexes. There is no canonical place where a domain can declare who they are, what they represent, or which resources are authoritative.

The proposed solution suggests adding a JSON-LD document to every website to solve the problem.

Humans struggle with domain authority as well. People have trouble telling what websites are legitimate and which aren’t. If we’re solving this problem for AI, perhaps we can find a solution that works for end users too.

And let’s set aside for the moment the irony that AI is supposed to replace all of our jobs and become a super intelligence and at the same time we also need to add special AI training wheels for it to use the web.

I’d like to see the Safari team’s response to the Web MCP API canonized as a W3C Design Principle. Maybe a slight rewrite like this:

When a site’s actions are hard for an AI agent to use, that is a gap in the page’s own semantics, and we should first seek to close it in the platform’s shared layers (HTML and ARIA), where the user, assistive technology, and agents all benefit. 

We can put it right next to the W3C’s Priority of Constituencies:

If a trade-off needs to be made, always put user needs above all…

User needs come before the needs of web page authors, which come before the needs of user agent implementors, which come before the needs of specification writers, which come before theoretical purity.

You’ll notice AI agents aren’t in that priority of constituencies, but assuming agents stick around, it might be worth considering where we’d put them.

I hope we can all agree that user needs come before agent needs. We should keep this in mind as we consider proposals to modify the web to support AI.