What can it do at three in the morning with nobody watching?
Put it to the vendor the way you would put it to a new employee handed a door key. Not can you be trusted, which invites a promise, but what do you do here when the building is empty. Every agent worth buying has some part that acts without being asked, and your exposure sits in the eight hours you are asleep rather than in the demonstration.
What the occasion actually opens
Something checks a list on an interval and starts whatever has fallen due: a nightly review, a chaser after a week of silence, the brief at eight. In the reference design that part wakes every thirty seconds, a setting on the Register, the standing page of every setting in ArkOne’s reference design for an executive agent.
Two properties decide what the overnight hours are worth to you. The first is scope: a schedule adds an occasion and no new power, so an unattended run may do what a conversation may do, checked by the same outbound rules. The second is the count. Taking a job and marking it taken happen in one indivisible step, so a job fires once even where a second copy of the scheduler is running, which is why the design keeps one and treats the extra as waste rather than as the safeguard.
The limit belongs in the same paragraph. Once means at most once. A scheduler that has stopped fires nothing at all and complains to nobody, because the part that would complain is the part that is down.
One job, two ways of counting it
Suppose the renewal invented for this work has sat a week without a reply, and a second copy of the scheduler was started by an overlapping deployment that evening. Both columns describe the same night, and both designs are working exactly as built.
| Through the night | Taking the job is one step | Seeing and taking are separate |
|---|---|---|
| The job falls due at three | Claimed once, held | Seen by both copies |
| A second later | The other copy finds it taken and stops | Both copies begin drafting |
| What the room drafts | One chaser | Two chasers, a second apart |
| What the outbound rules see | One request, held until working hours | Two requests, each held until working hours |
| What the client receives at nine | One message | Two identical messages |
| What anybody notices that night | Nothing to notice | Nothing until the client replies |
Exhibit 1. Illustrative. One overnight job, counted through a scheduler that claims work and one that does not.
Sit with the last row. Neither column produces an error or an alert anywhere. One of them is a system doing its work twice as diligently, and the first person to learn of it is your client.
The fourth row rewards a second look for the opposite reason. The outbound rules worked in both columns and held the message until morning. They were never asked to judge whether the same message had already been drafted elsewhere in the same building, and a guard at the door catches what leaves rather than what was duplicated upstream of it.
Pressing on the second half
A good answer arrives as a list. These are the scheduled jobs, this is what each may touch, and a scheduled run passes the same send rules as anything a person triggers. Then, unprompted, the sentence about counting: taking the job is one write, and here is where two copies were started on purpose to prove it.
An ordinary answer names the trigger and stops. You can schedule any task, or it runs overnight so your morning brief is ready. Both describe the occasion, neither the scope, and the two facts you want are what a task may reach and what stops it happening twice. Press hardest on the reply that they only ever run one copy: that describes this month’s deployment rather than the code, and stops being true the first time a restart overlaps.
Then the quieter half, which nobody volunteers. Who notices if the scheduler stops? A duplicate embarrasses you for a day. A schedule that stopped in March is found in June, as a quarter of chasers never sent to clients never followed up. Ask what watches it, and where that alarm arrives.
Overnight work is most of the value of an agent worth having, and none of it is safe by virtue of being unattended. Two sentences settle it: what the jobs may touch, and what makes each fire exactly once. The question about two copies running at once is the same mechanism seen from the other side.
component: answer-card
Asked plainly
What can an AI agent do overnight without anyone approving it?
Whatever its scheduled jobs are set to do, under whatever rules the design applies to a run nobody started. A scheduler is a trigger and not a permission, so the honest answer is a list of the jobs plus a statement of which rules a scheduled run passes through. If overnight runs skip the checks that a daytime conversation passes, that is the fact worth having before signing.
Why would an AI agent send the same overnight message twice?
Because two copies of the scheduler both saw the job fall due and both ran it. The defence is that taking a job and marking it taken happen as one indivisible step, so the second copy finds nothing to take. Where seeing and taking are separate steps, a gap opens between them, and two hands can reach through it a second apart.
How would I know if an AI agent's scheduled jobs had stopped running?
Only if something outside the scheduler is watching it. A scheduler that has stopped fires nothing and raises no alarm, because the part that would raise the alarm is the part that is down. Duplicate messages are noticed within a day; a schedule that quietly stopped can go a quarter unnoticed, so ask what checks that last night's jobs actually ran.