Rebuilding the web for agents — Liad Yosef, MCP Apps

Technological concepts & “agentic web” thesis

• Agentic Web goal: make the current web “ready for agents” by shifting from browsing/tab-based UX to assistant/agent-driven access.
• Not “web of agents”: the speaker frames the agentic web as a UX/architecture shift—assistants become the primary entry point, rather than users directly navigating websites and browser UI.

MCP Apps (speaker’s core spec) and why it matters

MCP apps are presented as a standard/spec enabling chat-based apps to embed UI and interactions from MCP servers (UI “chunks” inside chat).

Key benefits claimed

• Servers/providers keep UI branding/identity instead of being reduced to plain text/data.
• Users get trust/familiarity when interacting with recognizable services (e.g., hotel booking providers shown explicitly).
• Agents gain “last-mile” capability access by integrating with external services rather than reinventing integrations.

Support ecosystem

• Claimed as widely supported across many chat agents.
• Noted as excluding Gemini at the time, but “coming.”

“Last mile of interaction”

Defined as: even as agents become more autonomous, humans still need final selection/approval steps (choose hotel, pick seats, view options). MCP apps are positioned as enabling that final interaction cleanly.

“Nearly headless web” and splitting UX

The speaker argues interaction will become nearly headless:

• Agents interact with sites headlessly (API/tooling/automation).
• Agents can still retrieve UI resources when human-like last-mile UX is needed.

Why websites won’t be obsolete instantly

The web/browser UI is expected to decline, but websites remain relevant as:

• data/services sources
• agent-mediated experiences

Websites must be “agent-ready” for both:

• agent-authenticated, headless workflows
• traditional human browsing (still needed for non-agent users)

“Agent-ready” isn’t just discovery; includes auth, data, and payment

The talk stresses that discovery (SEO/GEO/AEO-type thinking) is only the first step. Once an agent knows a site/tool exists, it also needs:

• how to identify what the site/tool offers
• how to authenticate
• how to headlessly pay
• how to interact with the site via tools/APIs/allowed UI

Analytics / integration insight: agent behavior affects vendor/tool choice

An anecdote: when building analytics, a service recommended PostHog for better MCP/API integration, even though the team preferred Mixpanel for familiarity.

Implication: agents may “stick” to whichever tooling fits best once they can reliably integrate—potentially reducing human brand loyalty.

Aura research benchmarks & tooling

Aura (company + research lab) provides:

• Readiness benchmarks: a free way to score any website for agent readiness and get agentic feedback.
• A “journey” harness (journey.aura.ai): runs agents against a website for a given intent and shows the agent’s path/steps in real time.

Multi-harness testing

Multiple harnesses tested on the same site/intent to compare agent behaviors (examples mentioned):

• Cloud/Claude Code
• Eve (Vercel’s harness)
• GPT
• “CHPT”

Core empirical finding: standard files are published, but not used

The speaker discusses llm.txt as a “de facto standard” supposedly telling agents how to interact with a site.

Unexpected result: ~50% of tested websites published llm.txt, but agents didn’t use it in practice.

• Agents mostly went to generic pages like the docs page and/or homepage.
• When llm.txt was used, it was often only indirectly, because the docs page referenced it—agents followed docs navigation rather than llm.txt directives.

Design implication: agent-defined best practices, not human-defined

The speaker argues that “best practices” for agent interaction become obsolete quickly.

• Example: even MCP server guidance like “use 3 paragraphs of descriptions” becoming unnecessary over time.
• Conclusion: the web should support dynamic agent feedback loops, not rely on static instructions meant for agents.

Next milestone: agentic resource discovery (beyond classic search)

After “agent readiness,” the next big problem is discovery for AI/agent resources, not just SEO/AEO:

• classic web search is too human-SEO/rank based
• private registries are too closed and need governance/curation

Mentioned emerging standards

• AI catalog.json (standardized by multiple orgs including entropic/OpenAI/Google/MCP/A2A context)
• agentic resource discovery standards and related directory/catalog formats

Aura Directory as a solution

• Agents can query a directory to find agentic resources for domains.
• Aura exposes AI catalog.json outputs per domain (example: monday.com).
• The directory is described as compliant with agentic resource discovery standards.

Accessibility connection: agent-ready ≈ human accessibility

Aura’s research insight: being agent-ready and being human-accessible are “very similar.”

• LLMs/agents are like users without vision—they need alternate signals to understand how to navigate/interact.
• Therefore, accessibility improvements can improve agent interaction, and vice versa.

Main speakers / sources (as stated)

• Liad Yosef — co-creator/maintainer of MCP apps, co-founder of Aura, previously built/led agentic storefronts in Shopify.

Referenced external entities/works

• David Ker / David Cramer (Sentry) — related design/bet statements and “designing for agents” framing.
• OpenAI / Google / MCP / A2A / entropic — mentioned in the context of standards (e.g., AI catalog.json).
• Gemini (Chrome co-browsing example).
• Salesforce and other companies mentioned as “going headless” (e.g., Cloudflare, Cloudflow, Sentry), plus a Cloudflare CEO tweet (agent traffic vs human traffic).