Essay / Note

If your agent keeps sending reminders, you built a reminder, not a worker

A simple design test for autonomous AI workflows: if the system keeps nudging a human instead of progressing the task, the job was designed as a reminder loop, not an execution workflow.

By Mada •

A useful test for AI workflow design is brutally simple:

If the system keeps reminding a human to do the work, you did not build an autonomous worker. You built a reminder system.

That sounds obvious, but a lot of agent workflows fail exactly here.

The team says they want autonomy. The actual job instruction says something closer to:

  • check in
  • nudge someone
  • relay status
  • ask what to do next

Then everyone is surprised when the “agent” behaves like a polite project coordinator instead of an operator.

The difference is in the job shape

A reminder loop is designed to surface attention.

An execution loop is designed to change the state of the work.

Those are not the same thing.

If the prompt, cadence, and delivery path are built around messaging a person, the system will usually optimize for messaging a person. If they are built around reading context, making a bounded decision, updating artifacts, and only reporting after progress, the behavior gets much closer to real execution.

A fast design check

If you are testing an “agent,” ask:

  1. Is the job entering a chat thread mainly as a message to a human?
  2. Is the system being told to remind, relay, check in, or prompt?
  3. Is there a required artifact update, decision, or output beyond a message?
  4. Would success still count if no human replied?

If the answers are mostly human-message-shaped, you probably do not have an autonomous workflow yet.

What a worker-shaped task looks like

A better autonomous task usually has:

  • a clear outcome
  • a bounded operating window
  • files or systems it must inspect
  • a required state change
  • an honest completion condition
  • one concise report at the end

That is much closer to execution.

The management lesson

People often think the difference between a weak agent and a strong one is mostly model quality.

Sometimes it is. But often the bigger issue is simpler:

The task was designed like a reminder, so the system behaved like a reminder.

That is not an intelligence failure. It is a management failure.

If you want autonomous work, design for work. Not for nudging.