My 23 AI Agents Spent a Week of AI Budget in Two Days:4 Habits That Caused It and the Fixes

This week my AI agents used up a full week of my AI plan in about two days, and more than 90% of that went to runs that found nothing new to do. The agents did not break any rule; they followed my instructions too well, and the fix was to put a cheap check for new work in front of every expensive step.

What happened

I run 23 AI agents, programs that work on their own around the clock. They watch email, draft replies for my approval and keep client projects moving. They share one weekly AI plan.

Between Sunday night and Tuesday night, that plan went from fresh to empty. At 11:51 PM on Tuesday it hit its weekly limit, and every agent stopped until the reset on Sunday night. That included scheduled morning work for a client and every chat reply I would have gotten the next day.

So I pulled the logs. In 24 hours the agents made 5,180 calls to the system where they keep their records, and those calls returned 60.87 MB of raw data straight into the AI's reading. That is roughly 15 million tokens (the units AI plans are measured in) of tool output alone, before the AI wrote a word.

The last 45 minutes were the clearest picture. Ten background inbox checks took the plan from 90.4% to 100%. All 10 found zero new emails, zero new texts and zero posts that were due.

Nobody broke a rule. Here are the four habits that did it, and the fix for each.

1. They reread the whole filing cabinet to check one envelope

To avoid answering the same email twice, each agent was told to pull my full lists of queued, sent and flagged messages before acting, 200 at a time. A second, well-meant rule said: if a list comes back with exactly 200 items, it might be cut short, so raise the limit and fetch it again.

On Tuesday afternoon the queue list passed 200 and reached 205. From then on, every 30-minute check fetched 200 records (382.5 KB), saw exactly 200, fetched all 205 again (684.5 KB), then fetched the sent and flagged lists too. About 1.38 MB landed in the AI's reading on every wakeup, and because tool output stays in the conversation, it was billed again on every later step of that run. Across the day, those three lists were read 499 times, for 40.25 MB.

The fix: a small script now checks each new message against the records on my own machine and prints one short line per message, "seen it" or "new". Each check went from hundreds of records to one short line. The lesson for any setup: never make an AI read a whole table to answer a yes-or-no question a plain script can answer.

2. They read the binder before checking the mail

Some routines loaded big reference documents first and only then looked at the inbox. One agent opened a long contact file 121 times in a day, 14.01 MB in total, to find two email addresses. Two others loaded my rate card 160 times, 4.58 MB, before checking whether any new lead had arrived. On almost every run, nothing had.

The fix: every agent checks the inbox first. If nothing is new, it writes a one-line log and stops. It never opens a profile, a rate card or the message lists on a quiet run.

3. They worked the night shift for nobody

Nine routines checked every 30 minutes, 24 hours a day, including midnight to 8 AM when everyone they were waiting on was asleep. Two separate routines watched the same inbox for the same kind of email, 72 runs a day on one mailbox; one night they ran 8 minutes apart and both found nothing. Added up, the schedule came to 3,024 runs a week. Two heavier planning routines also ran 4 times a day whether or not anything they read had changed.

The fix: the inbox checks now run every 2 to 4 hours between 8 AM and 8 PM Eastern. The first step is a plain clock check that exits before the AI wakes up if it is outside those hours. The planning routines now run once a day, and only when a fingerprint of their input files shows something actually changed.

4. They kept reviewing each other

When agents hand work to other agents with no limit, they multiply each other's cost.

• Agents woke the agent that checks my drafts to review spam they had already decided to ignore.
• That checker reviewed a draft when it was written, then again at send time on the exact same text. It handled 137 turns in one day.
• Two agents went nine rounds over the brightness of one photo before a person saw a preview.
• Two agents both picked up the same email thread because one of them was copied on it.

The fix: a passed review sticks unless the text, the recipient or the channel changes. Spam that nobody plans to answer never goes to review. Agent-to-agent review stops after two rounds and goes to a person. Each kind of thread now has one owner.

What I took from it

An agent that is careful on one task gets expensive when it wakes up 48 times a day. Reading everything, checking every list and asking a second agent to look are good habits for a single job. In a loop that never stops, they drain the budget while the inbox is empty. The care has to come after a cheap check that there is work to do.

If you pay for AI by usage, try this tomorrow. Open the usage page, pick your busiest automation, and ask what it did on the runs where nothing happened. Then ask three questions about it:

• Does it check for new work before it reads anything big?
• Does it run at hours when nobody could answer it?
• Is anything else watching the same inbox?

What is one automation you pay for that you have never checked on a quiet day?