Loud. A channel where the agent posts what it finds on the sources you asked it to watch was useful for a fortnight. Then a competitor rebuilt a page, the source began reporting the same change on every read, and the channel filled with one fact wearing slightly different words. By the afternoon somebody had muted it. Nobody announced that or wrote it down, so the agent went on watching a list that no longer reached anybody, including you.
What went wrong
A watch has two quantities to bound, and they are not the same quantity.
The first is how many events it looks at. A source is read on a cadence, each read is compared with the last, and every difference is an event. That count follows the sources rather than the news, so a page republishing itself produces events at whatever rate it republishes.
The second is how many distinct alarms it may raise. One underlying fact, seen on forty consecutive reads, is one thing that happened and forty events. A design that bounds only the reading raises a fresh alarm on each read, because each read honestly found a difference and nothing has the job of recognising a fact already reported. The alarm is the output people live with, and it was never rationed.
The Cabinet, ArkOne’s reference design for an executive agent, bounds both. The Watch, the watchlists the room keeps on sources outside your company, evaluates at most sixty inbound events a minute, whatever the sources produce, so the reading is rationed by design. Separately, every signal carries a stable duplicate key built from what the signal is about, so scanning the same source twice cannot raise the same alarm twice. Those values sit on the Register, the page stating every setting of the design, and the twenty-fourth paper walks a night on one source.
The family resemblance is worth naming, because it shows the two bounds are one idea. The Door, the rules that decide what may leave the room, holds a limit of five messages to one person in an hour and refuses anything near-identical to a message already sent to that person inside a six-hour window. A rate and a duplicate rule, outbound this time, for the same reason: a recipient is a finite thing. The sixteenth paper sets out the Door, and the same absence on that side produced a message that arrived twice in eight seconds.
The record
Suppose a specialist insurer of three hundred people, watching four competitors’ published rates and one regulator’s notices. The insurer, its watchlist and the afternoon below are invented for the exercise.
| What the source did | Watch with neither bound | Watch with both |
|---|---|---|
| A rate page rebuilt once, then reporting the change on every read | An alarm per read, all afternoon | The first sighting raises one alarm; every later read folds into it |
| Comparison feeds echoing the same change | Their own alarms, worded differently | Same key, same signal, no new alarm |
| A burst of unrelated edits on the same page | Read in full, each one evaluated | Reading held to the ration; the surplus is read late |
| What the reader had by six | A muted channel | One alarm, and a regulator’s notice still legible beneath it |
Exhibit 1. Illustrative. One noisy afternoon on one watched source, with the two bounds absent and present.
Both watches saw the same afternoon. What differs is that one fact arrived once, and the surplus reading was postponed rather than allowed to consume the hour.
The damage is the muted channel rather than the noise. A watch that the people it serves have switched off leaves nothing in the software recording a change of state: the list is still configured, the schedule still fires, the log still shows healthy runs. What has gone is the only load-bearing part, which is somebody reading it. The next alarm that genuinely wants a person will be raised correctly, delivered correctly, and land where nobody opens it.
A ration means a genuine burst is sampled rather than seen in full: the events beyond the minute’s allowance are read late, so a real flood arrives in parts. The design chooses a legible sample over an unreadable flood, because a reader who can still read is worth more than a complete record nobody opens. Both are losses. Only one of them is one you can notice.
What happens
The fortieth scan of the same change folds into the first signal, and the channel carries one alarm the reader can still act on.
What your company sees
One alarm for one fact, in a channel somebody still reads.
What it means for you
Start inside your own company rather than with the vendor. Find out which alerting channels are currently muted, and by whom. A mute is a verdict on a watch that nobody filed as one, and it is the failure itself rather than a reaction to it: the noise is survivable, the silence afterwards is not. Every muted channel you find is a watch you are still paying for and no longer receiving.
Take the busiest day the watch has had and ask for two counts side by side: events evaluated, and alarms raised. A build that bounds both produces the two numbers and shows you the gap between them. A build that bounds only the reading hands you one number and calls it both. Turning the cadence down, which is what gets offered first, moves neither count in the direction you want, because a source read half as often still reports the same change twice.
One more question, and it is the one that gets vague answers: how does the agent know it has already told you something. A watch that can answer describes what it stores about a signal so that the same change, found again tomorrow, is recognised as the one you were told about today. A watch that cannot will say it filters duplicates, which is the claim rather than the method, and repetition is what you get.
Next: The department that acted on a mandate nobody had set.
Asked plainly
Why does an AI monitoring agent send the same alert over and over?
Because a source is usually read on a cadence, and a change that is still there on the next read looks like a new event on the next read. Unless a signal carries a key describing what it is about, so that the second sighting can be recognised as the first, every rescan produces a fresh alarm about a fact the reader already has.
How many alerts should an AI agent be allowed to raise?
Fewer than the number of events it sees, and the two numbers have to be bounded separately. A limit on how much the agent reads controls cost and attention. A limit on how many distinct alarms it may raise, and a rule that one underlying change raises one, controls whether a person can still read the channel. A design with only the first will repeat itself perfectly legally.
What is the real cost of an AI agent that alerts too often?
The channel going unread. Once people mute an alerting channel, the watch is switched off in the only sense that matters, and the next alarm that genuinely needed somebody lands somewhere nobody opens. The noise itself is a nuisance. The mute is the failure, and it does not appear anywhere in the software as a fault.