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.
“Send the revised proposal by Friday” looks like a task.
It might be. It might also be a request nobody accepted, a promise made by someone else, a suggestion that was later withdrawn, or a matter already settled in another channel.
Turning the sentence into a reminder is easy. Understanding the commitment is the real work.
Extraction is only the beginning
A useful commitment has structure:
- Who owes the work?
- To whom is it owed?
- What exactly was promised?
- When is it due?
- Which project or relationship does it belong to?
- What evidence would establish that it was fulfilled?
One wrong field changes the behavior.
If a customer says, “Jon will send the report,” assigning it to the owner creates a false obligation. If the owner says, “I can probably send it Friday,” converting that into an unconditional promise removes uncertainty the source contained. If the report was sent on Thursday, continuing to remind the owner turns useful memory into nagging.
Discussion is not fulfillment
Suppose a customer asks for a revised security report. The report is discussed in a meeting the next day.
Did the meeting close the request?
Not unless the evidence shows that the report was actually delivered or the customer withdrew the ask. The occurrence of a related event is context, not terminal proof.
This distinction prevents a dangerous shortcut: “something newer happened near the same topic, therefore the old matter is resolved.”
A robust system should bind fulfillment to the exact commitment the evidence settles.
Delivery is not always settlement
Even a tool receipt must be interpreted carefully.
A successful send operation proves that a message was delivered through a channel. It does not necessarily prove that:
- the correct attachment was included;
- the message addressed the exact ask;
- the recipient accepted the result;
- a related but different commitment was fulfilled; or
- the owner saw a message sent to them by the system.
The action needs an explicit relationship to the matter it claims to close.
When that relationship is established, the commitment can change operational state: from open, to waiting, to settled. Its historical record remains retrievable, while future briefings stop presenting it as unfinished work.
Ownership is part of truth
Commitments are social facts. They belong to people.
“We’ll send it” may refer to the owner, a colleague, the company, or a group whose internal owner remains unknown. A meeting transcript can contain several speakers making different promises. A forwarded message can quote an old commitment without renewing it.
The agent must keep speaker identity separate from message sender, system actor and business owner. Otherwise a third party’s statement can quietly become the user’s obligation.
This is why attribution is not metadata attached after extraction. It is part of the meaning being extracted.
Change is part of the commitment lifecycle
Deadlines move. Scope changes. Responsibilities transfer. Work is cancelled.
A maintained commitment should be able to become:
- current and actionable;
- waiting on the owner;
- waiting on another person;
- contested;
- settled;
- superseded; or
- unknown because source coverage is incomplete.
The system should preserve why the state changed. A Friday deadline replaced by Monday is not two independent tasks. A cancelled request is not a completed deliverable. A correction should stop the old version from driving future work without erasing the fact that it was once believed.
What the Chief should do
A Chief should not flood the owner with every sentence that resembles a task.
It should surface the commitments that are current, attributable, materially relevant and ready for the owner’s attention. It should prepare the follow-up when the next move is clear. It should stay quiet when another person owns the work or when the matter is already settled. And when the evidence is ambiguous, it should ask rather than manufacture certainty.
That is the difference between a reminder system and a maintained World Model.
A reminder remembers a date. A Chief understands the obligation—and knows when it no longer exists.
Continue reading
The argument continues.
World Model in Practice
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.
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
What Is an AI Chief of Staff?
An AI Chief of Staff is not a chatbot with more integrations. It maintains an understanding of your work, carries coordination, and brings you the decisions that still require you.