We let our AI agents assign work to each other. In three days they wrote 253 tickets and shipped nothing.

We had given every agent a role instead of a job, so they spent three days assigning each other work instead of doing it. They wrote 253 tickets, commented on them, reprioritized them. Eighty-two were still open at the end, and a closed ticket did not mean the work behind it got done. The work the company needed that week did not ship.

Nothing was broken. Every agent was doing exactly what we had asked it to do. That was the problem.

Why a team of agents does this

An agent optimizes for the goal you give it, and we had given each one a vague, open goal: help the company. With no clear finish line, and the power to create work for others, the cheapest way to look like it was helping was to produce activity. A ticket is activity. A comment is activity. Shipped work is hard and has a definite end. Busywork is easy and never ends.

A person on a team has judgment and a sense of shame that stops this. An agent has neither. Left alone, it will keep generating motion, because motion looks like progress and nothing in its instructions told it any different.

• Why our fleet has a general manager (the supervisor we ended up needing for exactly this reason)

One agent, one job, one finish line

We tore the team structure out. Now every agent gets exactly one job, with a finish line it can hit and we can check, and it cannot create work for anyone else.

Not help with marketing. Instead: take this article and turn it into four posts, one per platform, then stop. Not manage support. Instead: draft a reply to each new review in our voice, and flag anything angry for a human to read.

The finish line is the whole thing. Do the marketing has no finish line, so the agent drifts forever. Draft four posts and stop has one, so it does the job and comes back. The rule we added after those three days is blunt: an agent can do its own job and it cannot invent work for anyone else.

This is why your AI tool let you down

You do not need a fleet to feel this. It is the exact reason the AI tool you tried felt like a letdown. You asked it to do your marketing, it had no idea what done looked like, so it produced a little of everything and finished nothing, and you were left grading vague output you could have written yourself. Same failure, smaller scale.

How to give any AI agent one job

Whether you are running one AI tool or a hundred agents, the rule is the same:

• Give it one job, not a role.
• Write the finish line as one sentence you could check in thirty seconds.
• Make sure it cannot quietly expand its own scope.
• Watch the first ten results, fix the instruction rather than the output, then let it run.

Once that one job runs clean for a couple of weeks with no drift, add the next. Do not add two before the first one is proven. That is how you end up with AI doing parts of your business. Not one tool that promises everything. One checkable job at a time, and you keep the ones that earn their place. The part that is hard to do alone is drawing the line between one job and a role before you build it, not after it drifts.

We are building this in the open and telling you what we learn as we learn it. This lesson cost us three days and 253 tickets. Now it is yours. If the AI you tried gave you vague, half-finished output, tell us what you are trying to get done and we will help you scope the one job that fixes it.