Markdown source

Rebuilding the web for agents

Conference Context

Session Description

AI apps are the new browsers. And the web is not ready. For thirty years we built the web for human eyes, benchmarked by tools like Lighthouse: humans measuring human behavior. That era is ending. Bot traffic has overtaken human traffic, and we can't hand-write a benchmark for what comes next - every best practice goes stale the moment models improve. Your next customer isn't a human with a credit card - it's an agent with a protocol, and it would rather not see your interface at all. That shift moves the UX question from how a human experiences your product to how an agent does, and how a human experiences that agent. Already, some services report their MCP traffic outpacing their web UI. The agent is rapidly becoming the main surface, and it always takes the path of least friction. Claude Code might consistently prefer PostHog over Mixpanel simply because PostHog has the better agentic surface - and Mixpanel loses customers without a human ever weighing in. Meanwhile the agentic web protocol stack keeps multiplying, a new one seemingly every week. The harder problem isn't discovery - it's operability: whether the web can actually be run once an agent arrives, and what is the ideal stack for that. Should we lean into headless protocols, or ones like WebMCP that treat the UI as the source of truth? Does a site need to implement every new spec just to support every kind of agent? So we stopped guessing and watched real agents work the whole journey: finding, understanding, authenticating, acting, handing back to a human. The findings go against the last year of agent-readiness advice. Agents ignore the files we built for them, reaching for docs and homepages instead - and whatever they reach, they trust and act on. But when those files are linked properly, their usage jumps 4x. The format isn't the key for the agentic web. Reachability is. The web will never be completely headless. Some moments still demand a human: choosing a seat, comparing options, casually exploring. And agents aren't uniform - some want full headless access, others spin up a browser to fill the gaps, but that's a friction point, not a free fallback. So the web is going nearly headless, always with a human eye at the end. This talk maps the entire agent web landscape based on findings from real agent journeys research:

Media Evidence

MCP UI: Extending the frontier — Liad Yosef and Ido Salomon, MCP Apps (speaker-match related prior/adjacent AI Engineer video; captions: English auto-captions).

These are phone-photo slide captures from the Google Photos AIE Slides album. They are supporting slide evidence and do not override official schedule fields.

Evidence Graph

This evidence graph is generated from currently linked source material: official schedule text, related video pages, cached transcripts, visible slide text, dense/reconstructed slide pages, and AI slide-classification audits.

Media Signals

Agent Reading Notes

Use these signals to refine the synopsis, topic links, people/company context, and method notes. If a source is a related external video rather than an exact official recording, keep it framed as supporting evidence.

Transcript Status

Related video transcript availability: English auto-captions. The transcript has been fetched and converted to youtube o zkvb0iFDQ transcript. Treat this as supporting context, not a recording of this exact scheduled session unless later confirmed.

People

Notes

Supporting Slides

Slide Evidence

Synthesis

Synthesized Breakdown

This talk argues that AI apps are becoming the new browsers, so the web has to be designed for agents as operators, not only for human readers. It breaks agentic-web readiness into the whole journey of discovery, understanding, authentication, action, and human handoff. The key finding is reachability over format: agent-facing files and specs only help when agents can actually find them from trusted surfaces such as docs and homepages. The resulting design frame is a nearly headless web, where agents handle most execution while human-visible checkpoints remain available for judgment, comparison, and consent.

Speaker And Company Context

Topics Covered

Derived Links And Source Material

Novel Concepts / Clever Methods

Evidence Boundary

This synthesis uses the official schedule plus cached video transcripts. Official AI Engineer World's Fair San Francisco 2026 livestreams and cut videos are primary event video sources for transcript/slide evidence; external, historical, or speaker-matched videos remain supporting context unless manually verified as exact official event recordings.