What does it watch when nobody messages it, and what stops a noisy source from ringing all day?
There is a slide for this, usually late in the deck, with the word monitoring on it. When it appears, ask them to add a source in front of you. Name one of your own: a competitor’s price page, a regulator’s notices page. Then ask the second half of the question before anyone moves on, because adding a source is easy and the not-ringing is the part nobody demonstrates.
Between a change and your phone
Any agent can fetch a page overnight and notice it differs from yesterday. Every difference is an event, and most events are a banner, a footer date, a reordered menu.
The Cabinet, ArkOne’s reference design for an executive agent, keeps such sources on the Watch and puts three steps between an event and you. The first is triage, a judgement: does this touch a goal, a client or a recorded risk? An event touching nothing is filed and goes no further. The second is a duplicate key, and it runs in code: every signal carries a stable key built from what it is about, so a rescan of the same source and an echo of it on a comparison feed fold into the alarm already raised. The third is a ration, also code: at most sixty inbound events are evaluated a minute, so a page that goes noisy at three cannot buy the night’s whole attention. Both settings sit on the Register.
Counting down one noisy night
Suppose the source you named republishes itself all night, with feeds echoing it, in a scenario invented for the exercise. Ask the vendor to walk the count from the source to you.
| Stage | Ask them for | The shape of a real answer | Where it goes thin |
|---|---|---|---|
| What arrived | The number of events off that source overnight | A figure they can pull up | An estimate, offered as a range |
| What was judged relevant | What decided, and where it is written | A named test against goals and clients | Anything phrased as importance |
| What folded together | How repeats of one story become one | A key you can see on two signals | The model remembers what it flagged |
| What reached a person | The count, and which channel carried it | One item in a morning brief | It messages whoever needs to know |
Exhibit 1. Illustrative. One noisy source named in a demonstration, and the count at each stage between it and a person.
The third row is the one that decides whether you keep the feature switched on in month two. A story republished forty times overnight is not forty pieces of news, and the only durable way to say so is a key that is identical on the first fetch and the fortieth, whatever the wording changed to.
The fourth row is a different question wearing the same clothes. Deciding something is worth saying and being allowed to say it at four in the morning are separate mechanisms, and a vendor who answers the first as though it settled the second has one of them.
Show me the night it went wrong
A good answer names the sources you chose, then finds a night when one of them went noisy and counts it out: this many events, this many folded into one, this many alarms raised. That vendor has been woken by their own product and fixed it.
The reply to press on names no test at all: that the agent monitors the web for you, or that you can set up alerts for anything you like. Both describe the input rather than the filter. The follow-up is short: what decides that something matters, and is it a rule I can read? Where the answer is that the model judges it, you have a judgement made fresh every time, defensible for triage and no use as a ceiling.
Then one more, and it is the one that matters at four in the morning. When an alarm is raised, what stands between the deciding and the message leaving? A watchlist with a direct line to your staff has no rules about repetition, rate or the hour. Ask whether an alarm goes into a brief you read at eight, or into a notification tonight, and who chose which.
What the count is for
Every monitoring feature demonstrates well, because a demonstration lasts an hour and nothing is noisy for an hour. The failure arrives in week three, quietly, as a thing you stop reading. Ask for the counts now and you are asking whether somebody built for week three. The three steps and their limits are set out in the paper on the Watch; what happens to an alarm once it wants to leave is the earlier question about the gap before a message goes.
Asked plainly
What does an AI agent do when nobody is talking to it?
In a design built for it, a scheduler wakes on a cadence and fetches the sources you named: a competitor's price page, a regulator's notices, a news search on your largest accounts. Each fetch is compared with the last, and every difference is an event. Most events are nothing, so what matters is what stands between an event and a message to you.
Why does the same piece of news reach me several times from an AI agent?
Because each fetch produced a fresh event and nothing recognised them as the same story. The defence is a stable key built from what a signal is about, so a rescan of the same source, and an echo of it elsewhere, fold into the alarm already raised. Instructing the model to remember what it flagged is weaker, because a later run starts with no memory of the earlier one.
How do I stop an AI agent alerting me all day about a noisy source?
Ask for two things that run in code rather than in the model's judgement: a duplicate rule that folds repeats of the same signal into one, and a ceiling on how many events may be evaluated in a minute. The ceiling means a source that goes noisy is read late rather than allowed to buy the whole night's attention and bill.