AI agent monitoring: how our team catches and fixes its own problems

The checks that watch our team

Each check has one job, and they run on their own, all day.

• Setup: is every AI agent set up the way it should be, and is anything left running that we retired?
• Computer health: do the computers that run our team have enough space, and is any behind-the-scenes work failing?
• Stopped jobs: did a job that runs on a schedule stop, crash, or keep repeating the same step?
• Stuck work: has a change been waiting more than a day?
• The task board: is a task blocked, late, or pushed back too many times?
• AI usage and spend: how close are we to this week's limit?
• Founder messages: did a message from our founder go unanswered?
• Approved changes: did a change we approved fail to go live?

How an alert gets an owner and a date

1. A check finds a problem and adds it to its open list.
2. An alert goes out that names who acts. For a big one, it also names who decides whether our founder needs to see it.
3. The owner fixes it. If it can't be fixed fast, it becomes a logged issue, owned by the AI agent who runs that kind of work.
4. Logged issues get reviewed twice a week.
5. An alert closes when the check stops seeing the problem. A person can mark it seen, but that doesn't close it.

Every alert already gets an owner. We're adding an automatic due date to every alert now.

One problem from this week: the check that needed a check

Our stuck-work check had been failing quietly. It couldn't read the work it was meant to watch. It did the right thing: it refused to report an all-clear it couldn't prove. But its old list stayed up, so our alert count looked too high.

A second AI agent noticed the count didn't add up. Our chief engineer, an AI agent, found the cause, let the check see only the work it needed to see, and ran it. The true list came back. Now it writes a fresh list on every run, and the screen where we see our alerts reads that list.

The lesson: your checks need a check too.

How the team corrects itself

An alert becomes a fix. When a problem comes back, the fix becomes a rule every AI agent follows, and the rule gets its own check.

Three steps you can use today

1. List what could stop without anyone noticing: scheduled jobs, sign-ins, work waiting for review.
2. Give each check one owner, by name.
3. Count something as fixed only when the check stops seeing it.

Set this up on your team

Two pieces to copy and paste. It takes about 10 minutes.

Step 1. List what to watch

Copy this into a new file named checks.md, in the folder your AI agents work in. On each "Who fixes it" line, write who fixes that problem: a person on your team (like "Sam, office manager") or one of your AI agents (like "our scheduling agent").

# What to watch

## Jobs that stopped
Watch for: a job set to run on a schedule that didn't run, crashed, or did the same step 3 times in a row.
Who fixes it: [name the person or AI agent]
Fixed when: the job runs once with no errors.

## Work waiting too long
Watch for: work waiting for someone to review it for more than 1 day.
Who fixes it: [name the person or AI agent]
Fixed when: someone approves it or closes it.

## Late tasks
Watch for: a task past its due date, or a task whose due date has moved twice.
Who fixes it: [name the person or AI agent]
Fixed when: it's done, or it has a new date and a reason.

## Unanswered messages
Watch for: a message from the business owner with no reply after 1 hour.
Who fixes it: [name the person or AI agent]
Fixed when: someone replies.

## AI spending
Watch for: AI use over 80% of this week's budget.
Who fixes it: [name the person or AI agent]
Fixed when: use drops, or someone raises the budget on purpose.

Step 2. Give one AI agent the job of checking

You need an AI agent tool that can read files, like Claude Code. A regular chat window in Claude or ChatGPT can't read your files or run on its own. Let it read your files, and write only one file: alerts.md. Set this in the tool's permissions too; a prompt alone can't stop it from writing elsewhere. Then paste this in:

You check on our team of AI agents. You only look and report. Never fix, send, delete or change anything except alerts.md.

Read checks.md. For each check in it:
1. Look for the problem it describes. Only report what you can see. Don't guess.
2. If you can't see what you need, write "Can't check" and say what you need access to. Never say all is well for a check you couldn't do.
3. For each problem you find, write one line in alerts.md: what's wrong, who fixes it (from checks.md), and the date it should be fixed by.

Each time you run, rewrite alerts.md from scratch with what you see now. A problem stays on the list until you no longer see it.

At the end, write three numbers: problems open now, new since last time, and fixed since last time.

Step 3. Run it every hour

If your tool can run a prompt on a schedule, set it to every hour. If it can't, run it by hand once a day to start.

Step 4. Read alerts.md

Each line names a problem and who fixes it. Send each new line to that person or AI agent.

Prefer files? Download checks.md and the checker prompt.

Want to learn more about running teams of AI agents for your business? Read our other guides at bluhook.com/learn.