A robot is an agent with a body
Whoever authorised a robot is liable for it — which requires that the authorisation was recorded at the time, was narrower than the authoriser's own permissions, and can be revoked in a way that demonstrably stops the machine. Most robots deployed today record none of those three, so liability falls back to the site operator by default rather than by design.
The same three questions
Software agent governance asks three questions. Every one of them has an identical, unanswered form in robotics.
| The question | For a software agent | For a robot |
|---|---|---|
| Was authorisation recorded? | A grant issued at a moment, to a scope, by a named principal. | Nothing. A robot has the access its operator configured, with no record of who decided that or when. |
| Does revocation stop it? | Testable: revoke the grant, observe the agent halt. | Untestable in most deployments. Removing network access is not the same as removing authority, and a machine mid-task may not notice either. |
| Is there an audit trail? | Actions recorded against the grant that permitted them. | Telemetry, usually. Telemetry says what the machine did, not what it was permitted to do — which is the question after an incident. |
The three questions are the live half of the conformance suite published in @flashyos/aao, which asks them of any organisation running agents. They are reported as deferred rather than assumed, because an organisation that merely claims its agents are revocable has said nothing. The delegation principle they rest on — that authority may only ever narrow as it is passed on — is the least-privilege rule stated in NIST's security glossary, applied to a machine rather than to a process.
Why this is not a metaphor
The claim is not that robots are like agents. It is that the governance artefact is literally the same one.
An AAO charter names an organisation's standing responsibilities, the scope each may act within, the threshold above which a human must approve, and the human who answers. Nothing in that document mentions software. A charter that governs a research agent governs a forklift by changing the capability list — and the reason nobody has done it is that robotics and agent governance have been separate industries, not that the artefact does not fit.
Delegated authority that only ever attenuates — a child grant can never exceed its parent — is the property that makes a fleet safe. A robot holding broader access than the technician who deployed it is the failure mode nobody has designed against, and it is exactly the failure flashyID was built to make structurally impossible.