Every meeting in this series so far began because somebody spoke: a client wrote, a colleague asked, and the room woke to answer. One part of the room does not wait to be asked, and the question to put to it is the one you would put to a new employee with a door key. What do you do here at three in the morning, when nobody is watching?
Every thirty seconds, awake or not, something checks a list. That something is the Clock, the scheduler that acts unasked, one of the eleven parts of ArkOne’s reference design for an executive agent, the Cabinet. It looks for work that has fallen due and starts it. What it may do with that work is this paper’s subject, and the answer is shorter than the fear.
A trigger, and a claim
The Clock keeps a list of jobs, each with a time: a nightly review of open initiatives, a follow-up if no reply arrives in a week, the morning brief at eight. Every thirty seconds it checks for anything due. When it finds a job it opens a meeting, the same meeting a message would have opened: the Chair, the one voice that speaks for the room, reads the Brief, consults the Seats, searches the Library, then asks the Door, the rules that decide what may leave. The Clock adds a trigger and no power. Whatever the room could not do at three in the afternoon it cannot do at three in the morning, and the Door’s working-hours rule applies to a job as to a reply.
The second thing the Clock does is why it exists as a named part. Taking a job and marking it taken happen in one indivisible step, which the Register calls the claim. There is no moment between seeing a job due and holding it in which a second pair of hands could reach for it. The sentence to carry: a job fires once and only once.
That matters because a scheduled system embarrasses a company by doing the same thing twice. Two copies of a scheduler, started by accident or an overlapping deployment, both see the follow-up due and both send it, and the client gets two chasers a second apart. An indivisible claim closes that gap. The design also runs a single Clock, since the claim already guarantees once and a second would waste work racing the first.
Two limits sit beside that guarantee. Thirty seconds between wakings means a job starts within half a minute of its time, never on the second. And once means at most once: a Clock that has stopped fires nothing at all, without complaining, because the part that would raise the alarm is the part that is down. Firing twice is loud and firing never is quiet, so ask what watches the Clock.
One job, claimed once
Suppose the renewal that runs through this series, its client invented, has been open a week with no reply, and a second Clock started by mistake that evening. Two other parts appear below: the Minutes, the decisions written after each meeting and read back at the next, and the Record, the ledger of every action and its reason.
| Moment | The Clock | The room |
|---|---|---|
| Three in the morning | Wakes, as every thirty seconds; finds the nightly review due | Nothing yet |
| The same instant | Claims the job and marks it claimed, in one step | A meeting opens. No person spoke |
| A second later, the second Clock | Wakes; finds the job already claimed; does nothing | One meeting, never two |
| Inside the meeting | Holds the claim; starts nothing else | The Chair reads the Minutes: the renewal is open, a week without a reply. It drafts a chaser and asks to send |
| At the Door | Takes the rescheduled send onto its list for nine | Refused: the client is outside working hours. The reason is returned; the Chair schedules for nine |
| Nine in the morning | Wakes; finds the send due; claims it, once | The chaser leaves through the Door, once; the Record notes who sent it and why |
Exhibit 2. Illustrative. One job falls due at three in the morning: claimed once, run under the Door's rules, sent at nine.
The Clock gave the room a moment and nothing else. Everything the meeting did at three in the morning it could have done at three in the afternoon, and what it could not do it still could not. The chaser was drafted at night and left at nine because the Door’s working-hours rule, from the fourth paper, does not care what started the meeting.
The claim made once mean once. The second Clock in the third row is the failure that produces two chasers; it woke, looked, found the job gone. The same write that saw the job took it.
And the limit the fourth paper stated holds at night too. If the Door itself errors, the design lets the message through and writes the fault down, and at three in the morning nobody notices. A company that would rather its overnight payments and filings fail closed may set that class of tool to refuse.
Fifteen is the cap on the Register, not a target. Press run and count them.
A meeting has not been run yet. Press run to watch one count, fan out to the specialists, come back, and stop.
What this arms you to ask
Ask what the agent can do at three in the morning with nobody watching, then what happens if two copies run at once.
A good answer to the first lists the scheduled jobs, and says a scheduled run passes the same send rules as a conversation. A good answer to the second names the claim, says it is one indivisible step, and shows the test where they started two copies and one message left. Neither weak answer describes a rule. On the first question it is that you can schedule tasks, which names the trigger and says nothing about what a scheduled task may do. On the second it is that they only run one copy, a fact about today’s deployment that stops being true the first time a restart overlaps.
The follow-up is the same for both. Ask what in the code stops a scheduled job taking a path a conversation could not, and what stops the second copy. Then ask who would notice if the Clock itself stopped.
Next paper: What an agent may send, to whom, and when.
Asked plainly
What can an AI agent do overnight when nobody is watching?
Whatever its scheduled jobs say, under whatever rules the design applies to a scheduled run. In the reference design described here, a scheduled job opens an ordinary meeting: the same orchestrator, the same specialists, the same outbound guards. The scheduler adds a trigger, never a power, so a message drafted at three in the morning still waits for the recipient's working hours before it may leave.
Why did an AI agent send the same message twice?
One cause is two copies of the scheduler both seeing the same job fall due and both running it. The defence is a claim: taking the job and marking it taken happen in one indivisible step, so a second copy finds the job already claimed and does nothing. Without that step there is a gap between seeing and taking, and two hands can reach through it.
What happens if two copies of an AI agent run at once?
In a design with an indivisible claim, each job still fires once; the second copy wastes effort racing the first and wins nothing. In a design without one, the same job can run twice, and if the job sends a message, two messages leave. Ask the vendor what in the code prevents the second run, and whether they have ever started two copies on purpose to check.
Ask the vendor
The questions this part arms, each with the good answer and the answer that arrives instead.