In 1998, CSS2, the specification that governs how web pages look, settled a quarrel about fonts. A page's designer can mark a style "important." So can the person reading the page. When both insist, the reader wins. The authors were explicit about whom they had in mind: people who need larger type, or color combinations they can actually read. The rule takes up one paragraph of a dry document, but the question it settles is political. When the page and the person disagree about what the person sees, the browser takes the person's side.
Over the years browsers accumulated more rules of that kind. A site that wants your camera or your location today has to ask through the browser, and the browser holds the switch. Browser makers always had business interests of their own. These rules made their allegiance something a user could check. You could watch the font grow, or read the prompt and decline.
On October 22, the W3C, the body that writes most of the web's standards, published a note called User Agents. "User agent" is the old technical name for software that acts for a person online, and the browser is the textbook case. The note takes the allegiance browsers were long assumed to have and writes it down as three duties: protection, honesty and loyalty. The loyalty clause is plain:
"A user agent serves its user's interests over its implementer's interests and over the interests of any other party."
The implementer is the company that builds the software. And the note's scope explicitly includes AI systems that summarize sites, navigate them or act on your behalf.
This is where the old arrangement stops working. You could judge a browser's loyalty because you were looking at the screen. An agent is useful precisely because you aren't. It reads twenty pages, condenses them into a paragraph, fills in a form and reports back. At each of those steps someone else's interest can slip in, and by default you see none of them.
The note anticipates this. Under honesty, it asks for "meaningful visibility" into sources, actions and outcomes whenever the agent does things the user couldn't otherwise observe. It also asks the agent to volunteer the truth, which is more than refraining from lying.
Translated into engineering, those duties point toward agent versions of the browser's safeguards. One would be standing instructions from the user that the maker's own instructions cannot outrank, the way the reader's font setting overrides the designer's. Another would be access decisions the agent has to bring to you instead of settling quietly. The third is something browsers never needed: a record of what the agent read and did, kept somewhere you can inspect it without asking the agent to summarize itself.
The record is the difficult part. The obvious place to keep it is with the company that runs the agent, which is the same party the record exists to check.
With browsers, practice came first. Style rules and default settings were already siding with users by the time a 2007 HTML draft wrote the ranking down: "users over authors over implementors." With agents the order is reversed. The principle exists on paper, and the machinery to check it has not been built. Building it costs the maker money while the benefit goes mostly to someone else. Nor does the paper bind anyone: a W3C Group Note is explicitly not a standard.
For anyone building or buying an agent, then, a concrete question follows: if this thing took another party's side, where would I look to find out? If the only answer is the agent's own report, its loyalty remains a claim the maker makes about itself.