How Do Loops, Run-Until-Done, and Schedules Differ?
The three features look alike but serve different purposes. A loop repeats the same task at a fixed interval and fits work where the AI needs to keep picking up the in-progress job and pass it to the next stage, such as processing new rows in a sheet every 30 minutes.
A run-until-done feature, which keeps going until a condition is met, suits a single complex job with a lot to get through at once. A schedule (cron or routine) works even when your PC is off, doing its job at a set time and reporting back, closer to an AI employee. None of the three is always better; you pick by the task.
Which Work Fits a Loop, and Which Doesn't?
Loops fit two cases well: when work moves in stages and must always be handed to the next stage, and when the same work keeps piling up in real time and needs periodic checking.
Using a loop as a Monday-morning reminder, on the other hand, is wasteful. That kind of periodic alert belongs to a schedule. Think of a loop as a tool for when the AI must keep taking over work that is already in progress.
What Did We Actually Put on a Loop?
We bundled customer-feedback triage and outbound-lead email drafting into loops. Frequently arriving feedback was checked hourly; outbound leads that needed emails were checked every 30 minutes.
Instead of a person opening the sheet each time to check and instruct, the loop fills in the triage results and reply drafts, and even prepares the email drafts. Loops you run often are defined in a separate ledger document, and only the ones marked active are registered and run in one go.
Why Are Guardrails the Core?
When a loop runs, people monitor less. So the more an output leaves the building, the safer it is not to automate it. Even with email, we did not automate the sending, only the drafting, and left the final send to a human who reviews and clicks.
One more thing: a loop runs only while the session is alive on your PC and deactivates after at most seven days. Because you give it context and let it decide the next step, it also burns tokens faster. Rather than forcing every task into a loop, it is more realistic to set up just two or three real-time, repetitive jobs.