Your agents are driving the business. Who has the watch?
Watchbill™ puts every AI agent on a named watch, under a named human, on an append‑only record. When someone asks you to prove the supervision holds, hand them the watchbill.
Watchbill · one watchA real watch, redacted · 17 September 2026
Station
packaging and distribution
Agent
WBX27
Watch officer
Joseph AliseThe named human on this watch · holds the authority to stop it
Window
17 September 2026 · order placed 00:22 ET
Last proposal
a private, reproducible install package, built and staged for Joseph Alise
Last commit
a second script, run with a cleared environment; 8 of 8 checks passedRun by Joseph Alise
Last refusal
11:25 · a read-only command, over-blockedRefused before it ran · an over-block, kept on the record
It arrives as a question you cannot answer with adjectives
Two moments bring people here. Neither one is about how well the agents work.
01
The questionnaire
Someone with leverage is asking for evidence of AI supervision, and the honest answer today is a paragraph of adjectives.
02
The near miss
An agent did something nobody authorized, it got caught late, and the postmortem found there was no record of who was supposed to be watching.
Agents are standing watch on your production systems tonight. Watchbill is how you know who put them there, who is over them, and how you prove it when someone asks.
The standing orders
Three promises, and every artifact traces to one
01
No unmanned watch
Every agent stands a named watch, with a defined station, under a named human who holds the authority to stop it.
02
Nothing irreversible without the human on watch
Agents propose. Humans commit.
03
The record survives the watch
Every assignment, proposal, commit, and refusal lands on an append‑only record you can hand to an auditor, an insurer, or a customer without preparing a word.
The watch model
Delegation made explicit, bounded, and revocable
The watchbill puts a person’s name against every station.
Roles and authorities, at principle level
The answer to the asker
Show the watchbill.
Not a paragraph about how seriously you take AI safety. A roster: every agent on a named watch, under a named human, with the record of who held authority, what was proposed, what a person committed, and what was refused.
A passed check is a claim, not a clearance. The record shows what a human committed in a given window, and it shows what the gate refused. It does not promise that nothing can ever go wrong, because the product exists to keep people from making exactly that kind of promise.
