Steps
- Run the agent manually first — never schedule something you haven’t watched work.
- Set the cadence the workflow actually needs: weekly for digests and reports, daily for follow-ups, hourly for triage.
- Anchor it to a real moment — Monday 08:00 before standup, Friday 16:00 before the week-end review — not an arbitrary default.
- Check the first two scheduled outputs like a code review. If they’re right, the recipe holds.
- Re-check whenever you change a memory or a connector — the routine runs the current recipe, not the one you verified.
Scheduling anti-patterns
- Scheduling on day one — you haven’t seen the output yet. Earn the schedule with two good manual runs.
- Hourly because you can — a digest nobody reads hourly becomes noise that gets muted forever.
- Editing the prompt and forgetting the routine — the schedule fires the new prompt at 23:00 Friday. Test changes manually first.
Next: beyond one agent
When a single agent’s context gets crowded, multi-bot workflows split the work into a pipeline.