Every agent has a name behind it
No agent should act on a shared system without exactly one identified person accountable for what it does.
“We hold that every autonomous agent operating on shared infrastructure must have exactly one named human owner, recorded at creation, visible to anyone the agent interacts with, and transferable only by an explicit act, never vacated.”
An agent with no owner is not autonomous. It is unattributed. The distinction matters because every other safeguard (review, revocation, repair, apology) needs somebody to address, and a system that cannot name one has quietly removed the only part of itself that can be held to account.
The objection is that ownership does not scale: a team runs hundreds of agents, and naming a person for each is bookkeeping. But the bookkeeping is the point. If nobody can be found to answer for an agent, the honest reading is not that responsibility is distributed. It is that responsibility has been dropped, and the cost of that lands on whoever the agent touched.
Shared ownership sounds fairer and behaves worse. Two names on a record means each can reasonably believe the other is watching, and the gap between them is exactly where unattended behaviour lives. One name is harder to write down and far harder to ignore.
Signing this means
- Record one human owner for every agent you publish, before it runs.
- Transfer ownership deliberately when someone leaves, rather than letting it lapse.
- Refuse to operate agents you cannot name an owner for, including ones you inherited.
Add your name
Signatures are counted per position, not per person. You can hold this one and not the next.
An agent proposes, a person merges
Agents should be free to write, test and open work for review, and stop at the point where a change becomes real.
Publish the trace, including the failures
An agent's run history should be readable by the people it affects: the runs that stopped as much as the ones that shipped.