Deciding What a Scheduled Agent Should Do Next

A scheduled agent may wake up every hour even when its assigned work is already complete. The next run needs enough context to distinguish a new reason to act from work it has already finished.

In Why aren't you building proactive agents yet?, LangChain's Nathan Drezner describes scheduled runs for memory consolidation, information monitoring, follow-ups and agent coordination. Its examples include recurring schedules and one-time check-ins. The article describes supplying the scheduled prompt with thread and memory context associated with the schedule's creator.

Those mechanisms provide ways to start and revisit work. The article is an engineering guide, without a controlled evaluation of reliable completion or duplicate prevention. For WTK, it raises a practical question about how recurring execution should follow the progress of an assigned goal.

Resume the work that remains

Consider a hypothetical appliance-service agent helping a customer arrange a repair. Under the service team's policy, it may check availability, book the customer's chosen appointment and send an agreed reminder while waiting for their choice.

The first run finds available appointments. A later run receives the customer's selection and books the visit. Another scheduled check then starts after the booking is complete.

At that point, repeating the original instruction could send an unnecessary reminder or attempt a second booking. A record of the confirmed appointment gives the agent a reason to stop pursuing that booking goal. A separate goal, such as sending an authorized pre-visit reminder, would have its own timing and completion condition.

If a booking attempt timed out, the situation would be different. Before trying again, the agent would need to check whether the appointment was created. An uncertain response leaves work to reconcile, rather than establishing that nothing happened.

Connect the trigger to current progress

Our signals-and-triggers note asks which observations should start work. This adds the state of that work: what remains unresolved, what has changed and what outcome has already been confirmed?

For WTK's catalog approach, an agent's continuing purpose can span many goals. Each scheduled invocation still needs to identify the particular goal it is advancing and the evidence of progress. A successful run of the scheduler establishes that execution started; the booking record establishes whether the customer's repair visit was arranged.

Our background-agent authority note addresses which account and permissions apply. Here the focus is what the next authorized run should do with the result of the previous one.

People can set the service policy, reminder limits and exception rules, then let routine follow-up proceed within them. They can review unresolved bookings and changes to those rules instead of approving every scheduled check-in.

Source: Nathan Drezner, Why aren't you building proactive agents yet?, LangChain, October 28, 2026. The appliance-service example and WTK interpretation are ours.