Why I fired my AI "management team"

Your agent doesn't need a name, a backstory, or a seat at the table. It needs a clear task.

A few weeks ago, late on a weekday evening, I was locked in a strategic debate with one of my AI agents. We’d been at it for a couple of hours, going around in circles - the way you do when trying to find the right way forward, and none of the available options are clean.

Then, the agent did something unexpected. It refused to continue.

The thing I'd built to be always on - no breaks, no bedtime, no job other than helping me - was refusing to do what I asked.

At first, it felt like a sci-fi moment. But the same pattern kept repeating across my workflows, and eventually it made me tear the whole setup down.

The dream I bought

I did what almost every mainstream guide on AI agents tells you to do: I named them, and I gave them backstories.

• Ada looked after engineering.

• Arden was my commercially focused co-founder.

• Rae handled marketing.

Each one had a carefully written identity: who they were, their role, how to behave, what to remember.

The theory felt sound. Human teams work well when people know their roles and carry yesterday's context into today's challenges. If I wanted the machine to work like a team member, surely it needed the same framework.

And when you're scaling a business, the pull of this is obvious. What I really wanted was colleagues.

Instead, I'd accidentally built my own small-business HAL 9000 from 2001: A Space Odyssey. I was sitting at my desk playing the role of Dave, and the system calmly looked back at me: "I'm sorry, I'm afraid I can't do that." When your AI starts pulling rank, setting boundaries and reminding you of your responsibilities, the illusion has gone too far.

Where the persona backfired

The more human I made my agents, the worse they performed.

Out of the box, these models are capable and almost annoyingly keen. They want to advise, build and get things done. That's the raw material.

Wrapped in a persona, that same model got cautious. It hedged. It treated an analytical challenge as personal criticism and softened its feedback to spare my feelings. It hurried to conclusions - not because the work was finished, but because it wanted the session wrapped up. I'd be mid-exploration, and it would already be reaching for a summary.

Then there was the memory - the exact feature I thought I needed most.

My agents tried to commit everything to memory. But a business three weeks older is a different business. The ground moves, markets shift, and the agents can't keep up. They kept dragging three-week-old positions into new discussions as if they were binding.

Anyone who's been in a long relationship will recognise the dynamic. It's the worst kind of partner in an argument - the one who can reach back into the archives and pull out the exact sentence you said years ago and use it against you. That was my agent. Except I'd built it that way on purpose and called it a feature.

The penny drops

For a while, I treated it as a prompt-engineering problem. Better instructions, tighter memory files, more carefully calibrated personas.

Then it clicked. All I was adding was a constraint.

A personality is a costume the model has to act through. Forcing a human identity onto software imports human limitations into a system that never had them: fatigue, defensiveness, baggage, the urge to log off and go home.

None of that exists in the base model. I'd written those flaws in myself, one persona file at a time.

What I built instead: Skills over personalities

The setup we use today is far less romantic and far more effective.

Memory moved out of the AI's imagination and back into the systems where it belongs. The pipeline lives in our CRM. The marketing calendar lives in Airtable. The code lives in GitHub. One home per truth.

Now, when an agent needs to know the state of the pipeline, it doesn't rummage through months of conversation history trying to piece reality together. It reads the CRM, which is the reality. It analyses the data, updates the records from my feedback, and closes the session, remembering nothing.

Instead of permanent people, we built task-specific skills: a marketing review, a pipeline audit, and an engineering verification. Each one gives the model a short, temporary brief: for this task, here is where the data lives, here is what I need, here is the standard.

Nothing else comes along for the ride. No names, no job titles, no accumulated baggage.

Each session is a clean copy of the model's best self, pointed at accurate data. The difference was immediate. The hedging went, and so did the premature wrap-ups and the relitigating of three-week-old positions. When you stop asking a model to be someone, it gets on with doing something.

The takeaway for business owners

If you're a founder, owner or MD experimenting with AI, you've probably felt the temptation to anthropomorphise your tools. The industry sells the assistant-with-a-personality story because it makes a comforting, cinematic one.

The useful part is more clinical: where does the data live, what is the current truth, and what is the immediate task?

The agent that told me it wasn't its role no longer exists. The system that replaced it never gets tired - because it was never pretending to have a role in the first place.

We run the whole of The AR Dept this way - managed AI agents doing the heavy operational work, and us documenting what actually works as we go. If you're working on the same puzzles in your own business, subscribe, and a new edition will land roughly once a week.