The Safari team dropped a substantive critique last week, formally opposing the WebMCP API and articulating something the web platform community has struggled to pin down: how should we build standards for AI agents without abandoning the humans the web was built for? Mike Wyrzykowski authored Safari's position paper, and it's worth reading in full if you haven't already.

The Core Argument Against WebMCP

Safari's objection centers on a principle that should guide any platform-level decision: when a site's actions are difficult for an AI agent to use, that's not an agent problem—it's a semantic gap in the page itself. The fix belongs in shared layers like HTML and ARIA, where it benefits users, assistive technology, and agents equally. Wyrzykowski wrote: '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.' That's not just pragmatic—it's architecturally sound.

AI Agents as Assistive Technology

The Safari team goes deeper: an agent acting on a user's behalf is essentially assistive technology. It should operate sites the way users do, without sites singling it out for special treatment. The current WebMCP proposal explicitly states it's 'not designed for ingestion by accessibility technology'—meaning richer, actionable semantics flow to agents while screen-reader and keyboard-only users get a degraded experience. That's the fork in the road right there: are we building standards that lift everyone, or creating separate tracks?

Domain Authority Standards: Humans Struggle Too

The discussion isn't limited to WebMCP. A recent thread on the Web Incubator Community Group (WICG) mailing list proposed a standard to help AI understand domain authority—essentially giving domains a machine-readable way to declare who they are and which resources are authoritative. The problem is real: AI assistants regularly misidentify or misrepresent sources because there's no canonical identity layer. But here's where the author's argument lands squarely: humans also struggle with domain authority. We can't tell legitimate sites from convincing fakes either. If we're solving this for AI, we should solve it for people too.

Toward a W3C Design Principle

The author suggests canonizing Safari's response as an official W3C Design Principle—flipping the framing slightly: '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.' It would sit naturally alongside W3C's existing Priority of Constituencies: user needs come before author needs, which come before implementor needs, which come before spec writer needs, which come before theoretical purity. AI agents aren't in that hierarchy yet—but if they're sticking around, maybe it's time to decide where they belong.

Key Takeaways

  • WebMCP's architecture creates separate semantics tracks for agents versus assistive technology users—a fundamental accessibility problem
  • Safari argues AI agents acting on behalf of users ARE assistive technology and deserve equal semantic access through shared platform layers
  • Domain authority standards being discussed in WICG should benefit human users who also struggle to identify legitimate sources
  • The author proposes elevating Safari's argument to a formal W3C Design Principle alongside the Priority of Constituencies

The Bottom Line

The irony is thick: AI is supposed to replace our jobs and achieve super intelligence, yet we apparently need special training wheels for it to browse the web. Before we fragment the platform with agent-specific APIs, let's exhaust every option that improves semantics for everyone. Users come first—not because it's ideologically convenient, but because that's how you build a web worth using.