Truth Is Not Attention
An AI can know something is true and still be wrong to interrupt you with it. Reliable agents must decide support, relevance and attention separately.
A fact can be true, relevant to your business, and still not deserve a place at the top of your morning briefing.
This is where many proactive agents go wrong. They treat retrieval as selection: if the system found something related, it should say it. The result is technically grounded noise—an assistant that knows many true things and has no judgment about when to speak.
Truth and attention are different decisions.
Five questions, not one score
Before an item reaches a proactive surface, the system needs to answer several independent questions.
Is it supported?
What evidence or trusted belief backs the claim? Is it current, contested, settled or unknown?
Where did it come from?
Can the system identify the exact message, meeting, tool result or belief that supports it? Does the source have authority for this kind of claim?
How is it related?
What connects the item to the owner, company, project or current question? A name appearing in an unsolicited email is not the same as an established relationship.
Did the owner attend to it?
Did they reply, open the relevant thread, join the meeting, make a decision, or otherwise demonstrate attention? Transport delivery and human attention are not the same event.
Does it belong on this surface now?
A research answer, meeting brief, morning update and interruption each have a different job. The same fact can be appropriate in one and distracting in another.
Collapsing these questions into one relevance score makes the system easier to build and harder to trust.
The cold email at the top of the brief
Imagine an inbound sales pitch that describes itself as urgent and transformative. The message is real. Its claims are evidence of what the sender said. The sender may even mention topics related to the owner’s business.
None of that establishes that the owner has a relationship with the sender, attended to the message, or should begin the day with it.
If a briefing promotes the email merely because it contains strong language and relevant keywords, the memory may be correct while the behavior is wrong.
The system did not hallucinate the email. It hallucinated its importance.
Proactive systems need a higher bar
When a user asks an explicit question, broad retrieval can be useful. The user has declared the topic and invited an answer.
Proactive communication is different. The system chooses the topic, timing and interruption. That additional power requires stronger evidence of usefulness.
A proactive item should usually have a reason such as:
- a current commitment is approaching a meaningful deadline;
- a material belief changed;
- a meeting is imminent and preparation would change the outcome;
- a relevant person is waiting on the owner;
- source evidence conflicts and a decision is needed; or
- a permitted action can safely remove work.
“This appeared in a connected source” is not enough.
Silence can be the correct behavior
Good understanding creates restraint.
A settled commitment should remain in history without resurfacing as open work. A true but low-value fact may be available on request without entering a briefing. An item with incomplete source coverage may need qualified language rather than a definitive alert. Private context may be relevant to the system and still forbidden on the current channel.
The best outcome is sometimes no message at all.
This is not a failure to use the available data. It is evidence that the system understands the difference between possessing information and deserving the user’s attention.
A Chief manages the threshold
A human Chief of Staff does not forward every true thing to the principal. The role is partly a threshold: what can be handled, what can wait, what belongs in preparation, and what genuinely requires judgment now.
An AI Chief needs the same separation.
The World Model maintains what the system can responsibly believe. The Understanding Engine interprets the current situation. Attention policy decides what deserves to appear here, now. Channel and trust contracts determine what may be said or done.
These layers cooperate, but none should silently replace another.
The goal of a proactive agent is not to prove how much it found.
It is to make the owner feel that what reached them deserved to.
Continue reading
The argument continues.
World Model in Practice
A Reminder Is Not a Commitment
Reliable follow-through requires more than extracting a task. An agent must know who promised what, to whom, by when, and what would actually prove completion.
World Model in Practice
Your Agents Should Not Each Build a Second Brain
When every agent constructs its own memory of the business, truth fragments. A maintained World Model can give each permitted agent the same current understanding.
Agentic Understanding
Beyond the Knowledge Base: How a Company Learns to Understand Itself
Enterprise search can find what was said. A Company Brain must also know what is true now, what changed, who owns it, what deserves attention, and what may safely happen next.