Turning a repeated task into a scheduled run
Start with one agent, test it on its own, chain it into a workflow, then let it run without you.
The short answer: the path from “I do this by hand every week” to “this runs itself” is three small steps, not one big one — build an agent, test it alone, then schedule it.
Step one: build one agent that does one job well
Give it a name, a job, and pick which model it runs on. Keep its job narrow — “review this folder for typos” is a better first agent than “manage my whole project.” A focused agent is easier to trust, and easier to fix when it’s wrong. See how agents work.
Step two: try it before you rely on it
Run the agent on its own a few times, on real examples, before wiring it into anything bigger. This is the step that’s easy to skip and expensive to skip — an agent that’s subtly wrong is more costly once three other steps depend on its output.
Step three: chain it, if the job needs more than one agent
Some tasks are really a sequence — a reviewer finds problems, a fixer addresses them, a tester checks the fix. Workflows connect agents so one’s output becomes the next one’s input, and you can watch a run happen step by step instead of guessing what happened inside one long prompt.
Step four: schedule it
Once you trust it, give it a recurring time instead of a button you have to remember to press. Scheduled runs work through a queued list on their own — you can check progress anytime, including from your phone, instead of babysitting a run in real time.
Start smaller than feels necessary
The temptation is to design the whole pipeline up front. Don’t. Build the first agent, run it by hand a few times, and only add the next step once the one before it is boring — in the good sense, where it just works.
Try the schedule previewer — describe how often something should run and see the next few times it would fire.
Common questions
How many agents should a workflow have?
As few as the job needs. Two or three focused agents chained together usually beats one agent trying to do everything, but a workflow with ten steps is often a sign the job should be split into two workflows.
What happens if a scheduled run fails partway through?
You want to be able to see which step it got to and why — check the run’s progress rather than assuming either “done” or “nothing happened.”
Can I test a scheduled run without waiting for its actual time?
Yes — that’s what step two is for. Test the agent and the workflow by hand first; scheduling should be the last thing you add, not the first thing you build.