My AI agent runs too often and wastes money. How often should it check?

What happened

One of our monitors ran every 60 seconds. Each run started a new agent session. Most of those sessions found nothing new and ended.

We changed it to run every 30 minutes. That saved more than 1,250 sessions a day.

In the same cleanup, we found a social-content job running on an expensive model, and it was in an error state. It was costing money and not even working. We moved it to a cheaper model.

Why the AI did it

The agent didn't choose to run every minute. A person set that interval, probably because "check often" sounds safe. With a plain script, checking often is cheap. With an AI agent, each check can mean loading instructions and tools and calling a model, so every run has a real cost.

The question to ask is how fast you actually need to know. If no one will act on the answer within a minute, a one-minute check buys you nothing.

The same thing shows up in Salesforce agent setups: an agent that polls for new leads, case updates or sync status every minute, when the team only looks at the results a few times a day.

The fix, step by step

1. List every scheduled agent and its interval. Include monitors, digests and content jobs.
2. For each one, write down how fast someone acts on its output. Minutes, hours or the next morning.
3. Set the interval to match. Our monitor went from 60 seconds to 30 minutes.
4. Do the math before and after. Every 60 seconds is 1,440 runs a day. Every 30 minutes is 48.
5. Check each job's model and status. Look for jobs on expensive models, and for jobs stuck in an error state that still run. Fix or pause those first.
6. If you need instant alerts, use a trigger, not a poll. Have the system send an event when something changes instead of asking an agent to keep checking.

How to check it worked

• Count runs for a day before and after the change. The drop should match your math.
• Plant a real event, like a test record or a failed job, and time how long it takes to get flagged. It should be inside the new interval, and that delay should be fine for whoever acts on it.
• Check billing over the next few days. Spend on that monitor should fall with the run count.