The web is shifting from something we use to something that acts on our behalf. Agentic AI promises systems that read, decide, negotiate, and optimise for us, but when algorithms become our representative, a deeper question arises: who is it actually representing? Drawing on lessons the web has...

...already learned about mediation, semantics, and proxy-experience, this talk uses accessibility as a stress test for agentic systems. It explores where agents fracture first, and what that reveals about whether they are built to serve people, or merely optimise for statistical averages.