Rebuilding the web for agents — Liad Yosef, MCP Apps
Publishing agent instructions is not enough: tested agents mostly ignored llms.txt and navigated documentation directly, so agent readiness needs observed journey testing and discoverable interfaces.
Aura found that **almost 50% of tested sites** published llms.txt, yet almost none of the tested agents used it without another page pointing them there. Agents generally went first to documentation and then the homepage.
Test real tasks across several agent harnesses and inspect their paths instead of treating one metadata file as compliance. Expose callable resources through standards such as **MCP Apps** and **AI catalog.json**, while retaining UI for choices that still require human judgment.
Aura found that **almost 50% of tested sites** published llms.txt, yet almost none of the tested agents used it without another page pointing them there. Agents generally went first to documentation and then the homepage. Test real tasks across several agent harnesses and inspect their paths instead of treating one metadata file as compliance. Expose callable resources through standards such as **MCP Apps** and **AI catalog.json**, while retaining UI for choices that still require human judgment. The findings come from Aura’s own scans and journey tooling; the talk does not disclose the sampled sites, agents, or scoring method in enough detail for independent comparison. The broader prediction of a nearly headless web is directional, not established.
This shifts agent-ready web design from publishing a metadata file to testing actual discovery paths and exposing callable capabilities. It supports MCP Apps where choices need human judgment, but the undisclosed evaluation details make the llms.txt finding a prompt for journey testing rather than a general compliance rule.