All work

Assistant agent, always on

An AI agent works for a sole trader as project manager and secretary: reminders, planning, email and documentation, reachable via Telegram and email.

Status
In daily use
Technology
Open-source agent framework, Python, Telegram, email, Linux server

The situation

At first the agent could only handle recurring tasks. For one-off reminders it made up errors instead of setting the reminder, and it remembered nothing said in German.

The idea

What may go wrong is left to the model. What must not go wrong, reminders and sending email, runs through fixed, tested tools. The agent actively stores important facts in its memory.

How it is built

  • Reminders sit in a queue. A process checks it every minute; a lock prevents anything from running twice.
  • Delivery via Telegram, otherwise by email. Failed attempts are retried after 1, 5, 30 and 120 minutes, then an alert goes out.
  • The agent only sends email after explicit approval.
  • For sensitive data only models with a clear data policy are allowed; a model with unclear data retention was removed.
  • Credentials live outside the archive, and every action is logged.

Documented figures

  • 1 mininterval at which reminders are checked
  • 4retries after a failed delivery before an alert is raised

Status

In daily use (October 2026).

A similar problem?

Tell us in one sentence what it is about. We will tell you honestly whether we can help.

Write an email
Next projectSearch across our own archive

Assistant

Fixed answers, no AI