- Author
- Mark Nottingham
- Published
- 24 April 2026
- Kind
- Personal blog essay on web and protocol design, argument only
- Original
- www.mnot.net/blog/2026/04/24/agents_as_collective_bargains
- Subject
- Harnesses and agent loops
What it is
Mark Nottingham writes a personal blog about the Web, protocol design and Internet governance. This essay asks what AI agents would need to be trustworthy on a user's behalf.
What it says
- Users trust many unseen parties, and the essay lists cases such as TVs and cars.
- A browser negotiates with sites for all users, so everyone gets the same deal.
- AI agents today have no well-defined user agent role and no standards.
- Sandboxing alone is not bargaining, because agents asking for intrusive permissions would wear users down.
- Agents inside enterprises with contracts will probably thrive without such a role.
- Such a role will need constrained tools, possibly with permission models. The essay asks whether platforms or a public process will set it.
What evidence it gives
The essay offers argument and a list of linked news stories about privacy failures. It has no data or tests. The author calls parts of the account speculation.
Interests and limits
The essay is opinion by an author who writes on web standards. It does not test any design for an AI user agent. The author says a follow-up on legal protections is still to come.