What are scheduled tasks for an AI agent?

Ask a language model to remind you next Tuesday and it will agree, warmly, and then do nothing. Not because it is unreliable but because it does not exist between your messages. A model runs when called and stops when it answers.

For consumer chat that is fine. For a business where cases run for months, it is the whole problem.

Turn-taking is the default, and it is the limitation

A turn-taking agent only ever acts because someone spoke to it. Every useful thing that should happen in the gap — the document that never arrived, the quote that expires on Friday, the client who has gone quiet since March — depends on a human remembering.

The usual first attempt at fixing this is a heartbeat: wake the agent on a fixed tick and let it look around for anything to do. It works, but it pays for every tick whether or not there was work, and "roughly every fifteen minutes" is not the same as "at nine on the morning her police certificate is due".

What happens here instead

The agent schedules a specific instruction for a specific moment. That commitment is a row in the database — not something held in memory, in a session, or in a running process — so it survives restarts and deploys, and a scheduler executes it at the appointed time.

A task carries what it needs to be right rather than merely punctual:

• When, as an exact time, or a cron expression for something recurring, with an end date
• A time zone, so nine in the morning means nine where the customer is
• The instruction, encrypted at rest, describing what to do when it fires
• The channel it should go out on — web, Telegram or WhatsApp — because the follow-up should arrive where the conversation already lives
• Retries, because the moment matters and a transient failure should not silently eat it

Two kinds exist. The agent creates follow-ups itself as a consequence of the conversation ("chase the bank statements if they haven't arrived by the 14th"). Reminders are asked for outright.

Why this is a commercial feature, not a convenience

The work that makes money in a relationship business is almost all in the gaps. Nobody loses a case during the consultation; they lose it in the six weeks afterwards when a document did not arrive and nobody noticed.

An agent that can only respond is an agent that helps with the easy half. One that can commit to a moment months out — and be there — is doing the part your team currently does badly, not because they are careless, but because remembering forty open threads is not a thing people are good at.

What the agent does when it wakes

A follow-up is not a canned message. When the task fires, the agent runs with everything it normally has: the customer's space and what it remembers of them, the knowledge base, and its Skills. So "chase the outstanding documents" can check which ones actually arrived and mention only the rest.

It is delivered on the channel that customer uses — web, Telegram or WhatsApp — because a follow-up that lands somewhere they do not read is the same as no follow-up.