{ "video": { "id": "waI44NP1abk", "title": "Rebuilding the web for agents — Liad Yosef, MCP Apps", "duration": 1241, "upload_date": null, "channel": "AI Engineer", "source": "AI Engineer", "event_format": "In person" }, "analysis": { "video_id": "waI44NP1abk", "title": "Rebuilding the web for agents — Liad Yosef, MCP Apps", "one_liner": "Liad Yosef, co-creator of the MCP Apps spec, argues that assistants — not browsers — are becoming the entry point to the web, so sites must be built for agents first with a \"nearly headless\" last mile of UI, and shows Aura's data that agents ignore llms.txt and go straight to your docs page.", "summary": "Yosef frames MCP Apps — the spec behind chat-embedded UI in Claude, ChatGPT, Copilot and GitHub — as the piece that made the \"agentic web\" possible by letting providers ship UI chunks into a chat so they keep brand and identity instead of being reduced to a database. He argues websites and browsers as the source of truth are going away: users won't visit each company's agent, they'll use their own personal assistant, and browser/computer-use agents are a \"faster horses\" answer that forces agents through pagination and filters built for human perception limits. He points at Salesforce, Cloudflare and Sentry going headless, and at Sentry's David Cramer reversing his own bet against website obsolescence. Aura's crawls of the web found ~50% of tested sites publish llms.txt and agents essentially never used it — they went to the docs page then the homepage — which leads to his conclusion that agents, not humans, should define what agents need.", "key_points": [ "MCP Apps is the spec behind chat-embedded app UI; released a few months ago with Claude as the first client, now supported by every chat agent except Gemini (\"but that's coming\"). Everybody wins: providers keep brand and identity, users get trust and familiarity (you see it's Booking.com), hosts get capabilities they'd never build — OpenAI won't negotiate hotel deals or handle a venue seat change.", "The \"last mile of interaction\": as models become more autonomous, humans still need to pick the hotel, view a 3D model, choose a seat — so the web goes \"nearly headless\", agents interacting headlessly and bringing back UI resources only where a human decision is needed.", "Interfaces should be broken into atoms and composed by a personal assistant: \"I don't need 99% of bookings dashboard... I definitely don't need Jira's dashboard.\" Providers also skip integration work — Booking doesn't need to build a calendar integration because Claude already has that context.", "He rejects both browser agents (\"faster horses\" — an agent shouldn't be driven through filters, pagination and sorting perfected for human perception limits) and per-company agents; Google's WebMCP, where the site exposes JavaScript tools to Gemini in Chrome, is nice but simple — pushing a Salesforce dashboard through it is not what anyone wants.", "Evidence the shift is real: Salesforce went headless despite its UI being its main differentiator, plus Cloudflare and Sentry; Cloudflare's CEO tweeted agent traffic surpassed human traffic; David Cramer of Sentry, who a year ago offered to bet against websites being obsolete in 25 years, just published \"Designing for agents\" saying Sentry must design API-first as a surface.", "Agents have no brand loyalty: Aura asked Claude Code for the best analytics service, it recommended PostHog and kept insisting over Mixpanel — which they had used for a decade — because PostHog had better MCP and API. If your product forces an agent to spin up a browser, it remembers and won't come back.", "Aura's benchmark found almost 50% of tested websites publish llms.txt, yet the agents they ran never used it on their own — they went straight to the docs page, then the homepage; the 40% that did use it only found it because the docs pointed to the file. OpenAI has likewise stopped publishing tool best practices because models outgrow them — six months ago an MCP server wanted three-paragraph tool descriptions, now three lines suffice.", "Discovery is the next milestone and it isn't SEO/GEO/AEO: web search is based on human page rank, per-agent registries are closed and require submission, a central MCP registry has no curation or governance answer. Emerging standards are ai-catalog.json (Anthropic, OpenAI, Google, MCP, A2A) for how a site exposes itself, and Agentic Resource Discovery (ARD) for how a directory exposes itself — Aura's ARD-compliant directory generates ai-catalog.json per domain (e.g. Monday.com's MCP server and API server).", "Being agent-ready and being human-accessible converge: LLMs arrive at your site like users with vision disabilities — they don't see it and need other signals — so accessibility work helps agents and vice versa." ], "takeaways": [ "Design your product API-first as a surface and assume ~90% of the experience is the agent's experience of your site, not a human's — then add back only the last-mile UI (choosing a hotel, a seat, a 3D view) via MCP Apps.", "Be agent-ready for three audiences at once: your own agent, your customers' assistants (ChatGPT, Claude, Open Claw), and the human who still browses to your site directly.", "Don't stop at discovery/AEO — an agent that finds you still needs to know what you are, how to interact, how to authenticate, and how to pay you headlessly.", "Don't assume publishing llms.txt or agents.md makes you agent-ready; agents mostly hit your docs page and homepage. Test with real harnesses — Aura's free journey.aura.ai replays any intent on any site across Claude Code, Haiku, Vercel's Eve and ChatGPT, whose paths differ markedly — rather than guessing at best practices that go stale as models improve.", "Expose an ai-catalog.json and make your agentic resources ARD-discoverable, and treat web accessibility work as agent-readiness work." ], "topics": [ "agents", "mcp", "agentic-web", "agent-ux", "discovery", "headless", "accessibility", "standards" ], "tools": [ "MCP Apps", "MCP", "Claude", "Claude Code", "ChatGPT", "GitHub Copilot", "Gemini / Gemini in Chrome", "WebMCP (Google)", "Aura (aura.ai, journey.aura.ai, Aura directory, readiness benchmark + leaderboard)", "Shopify", "Sentry", "Salesforce", "Cloudflare", "Booking.com", "Airbnb", "AllTrails", "Jira", "PostHog", "Mixpanel", "Attio", "Eve (Vercel's harness)", "Haiku", "Open Claw", "Monday.com", "Vercel", "llms.txt", "agents.md", "ai-catalog.json", "Agentic Resource Discovery (ARD)", "A2A", "OpenAI", "Anthropic" ], "quotes": [ { "text": "I don't need 99% of bookings dashboard. I don't need 99% of Airbnb's dashboard. I definitely don't need Jira's dashboard.", "at": "05:11", "url": "https://www.youtube.com/watch?v=waI44NP1abk&t=311s" }, { "text": "that's a faster horses solution.", "at": "07:13", "url": "https://www.youtube.com/watch?v=waI44NP1abk&t=433s" }, { "text": "So we don't have brand loyalty, right?", "at": "12:17", "url": "https://www.youtube.com/watch?v=waI44NP1abk&t=737s" }, { "text": "But we don't want to rebuild the web for agents.", "at": "20:06", "url": "https://www.youtube.com/watch?v=waI44NP1abk&t=1206s" } ], "words": 4088 }, "summary_url": "/#waI44NP1abk", "transcript": { "html": "/transcripts/waI44NP1abk.html", "txt": "/transcripts/waI44NP1abk.txt", "vtt": "/transcripts/waI44NP1abk.vtt" } }