9 Women, 1 Month, 0 Babies: Why You Don't Need an Army of AI Agents

Anyone who needs 30 AI agents running at once can't prioritise. Why clarity matters more than the number of agents, and the six questions to start with.

Date: 4 October 2026 | Author: David Hefendehl

I was a PM for years, and now as then, 9 women still can't carry a baby to term in one month. Anyone who has to run 30 AI agents at the same time simply can't prioritise.

That might sound harsh, but that's exactly how I mean it. Before you head for the comments section, please hear me out for a moment, because I've got nothing against AI agents and use them myself every day. What bothers me is the idea that quantity automatically means progress.

Where the line about the 9 women comes from

The saying is older than any AI; Fred Brooks made it famous in 1975 in "The Mythical Man-Month". His observation from large software projects at IBM was that adding people to a late project only makes it finish even later. New people need onboarding and coordination, and every additional person creates new lines of communication that grow faster than the extra capacity gained.

Some tasks simply cannot be divided up, no matter how many hands you put on them. A baby takes nine months, a good campaign idea needs time to think it through, and a thorough data analysis needs someone who understands what the figures actually mean.

Every project manager knows this from their own painful experience. Yet we're currently acting as if AI has overturned this law, when in reality it has merely put it in new, shiny packaging.

Welcome to the agent arms race

And that brings us straight to the heart of the matter. Everything here on LinkedIn, on Insta, even on Facebook suggests that you need to have a gazillion agents running just to stay in the game.

You know those posts with a screenshot of a terminal showing twelve windows open at once. Underneath, there's something like "My agent team wrote 47 blog articles last night, qualified 300 leads and appealed my tax assessment", accompanied by a video of someone casually sipping coffee while the screens flicker in the background.

The big developer conferences of the AI providers do the rest when 27 bots are already running on stage first thing in the morning, just so that someone can relax and have a coffee. It's a good show, and that's exactly what a stage is meant to do: generate enthusiasm, sell products and win over developers.

It only gets silly when many of us take this show as the benchmark for our own Monday mornings. Then you sit there with your one task and feel left behind, because you're not orchestrating 30 agents and are "only" working with a chat window.

But let me tell you: you haven't been left behind. You're probably the most sensible person in the room right now.

What work really looks like for the vast majority of us

For the vast majority of us, the reality is that we spend hours on a single task. That could be research for a campaign, analysing customer feedback, sales forecasting for the next quarter or bookkeeping, just as much as the proposal for the big client or the presentation for senior management on Thursday.

None of these tasks can sensibly be broken into 30 parts and worked on in parallel. They follow a sequence in which you first understand, then structure, then implement and finally check, with each step building on the previous one.

Very few people will have to do all of this at the same time. And the few who actually do usually have a different problem than a lack of agents, namely one with focus, with delegation, or with a role that is simply cut too large.

On top of that comes a point many people underestimate: even 30 agents need someone to read, check, contextualise and approve their results. That someone is you, and you still have exactly one brain to process it all.

You're the bottleneck, and that's a good thing

Gloria Mark from the University of California has spent years studying how people work at a screen. One of her best-known observations is that after an interruption, it takes us on average more than 20 minutes to get fully back into the original task.

Now imagine you have ten agents running, and each one checks in at some point with a result, a question or an error. Each of these pings is an interruption, so you jump from campaign research to invoice checking, on to code review and then to social media planning. At the end of the day, you've seen a lot, but understood very little of it.

That has little to do with productivity; it's more like multitasking with better marketing.

More agents just shift the work around instead of taking it off your hands. Execution gets cheaper, while evaluation stays expensive. That's precisely why the value of AI isn't decided by the number of your agents, but by the quality of the decisions you make beforehand.

For us ordinary folk: prioritise

So for us ordinary folk, it's more true than ever that we have to prioritise. That sounds boring at first, I know, but it's the only skill that becomes more valuable rather than cheaper because of AI.

So before you fire up any agent, sit down for ten minutes and answer a few unspectacular questions, in writing and not just in your head on the side.

1. What's the most important thing that needs to get done?

This is where Eat the Frog helps, a method based on a saying often attributed to Mark Twain and made popular by Brian Tracy. If you tackle the "eat the frog" tasks first thing in the morning, the rest of the day can really only get better.

Put simply, the most important and often most unpleasant task comes first. Not the task you enjoy most, nor the one you've just found the coolest tool for, but the one that moves the needle most.

And this is exactly where the first trap lies, because AI makes easy tasks even easier. It's very tempting to have the agent knock out the five nice little jobs first, because that feels so wonderfully productive, while the frog is still sitting on your plate afterwards.

2. What's the most annoying?

This is the question about the low-hanging fruit, because annoying usually means repetitive, fiddly and boring. Typical examples are copying together data from three exports, turning meeting notes into tasks, or writing the twentieth version of the same email.

This is exactly where AI plays to its strengths, not because it's particularly spectacular, but because it gives you back energy for the important things. Every hour less of copy-paste is an extra hour you can use for thinking.

3. What's the most draining?

Draining is different from annoying, because draining tasks are the ones where you really have to think. They depend on judgement and on context that no AI in the world has.

This is where AI helps in a different way, as a sparring partner, as a counter-voice, or as someone who lays out a first structure for you that you can then tear to pieces. But the decision stays with you, because anyone who outsources it to an agent is outsourcing their responsibility, and that comes back at the latest in the next meeting with senior management.

4. What in the workflow can AI handle, and what do I need to check?

Now it gets concrete: you take your task and break it down into individual steps. For an analysis of customer feedback, it might look roughly like this:

  • Collect feedback from the various sources
  • Reliably remove duplicate and empty entries
  • Cluster the feedback sensibly by topic
  • Count how often each topic comes up
  • Interpret the most important patterns in the data
  • Derive concrete actions from the patterns
  • Prepare the results for management

AI is really good at the first four steps, and it can at least prepare the fifth well. The sixth and seventh steps, however, need you, because only you know which complaint is politically sensitive, which customer is about to cancel, and which topic sales has been blocking for months.

That leaves the second half of the question, namely what you need to double-check, and the answer is: a lot more than you think. Figures, quotes and everything that leaves the building, you always check, because AI results are drafts, not finished results. Anyone who forgets this usually finds out through a rather embarrassing email.

5. What should the output look like?

This question is almost always skipped, even though it may be the most important of all.

If you don't know what the result should look like, no AI can deliver it for you either. Should it be a table, continuous text, three bullet points for the board or a draft with sources, and how long, in what tone and for whom?

The more clearly you describe this, the better the result will be in the end, and that has precious little to do with rocket science. It's a good briefing, just like the one you'd give a working student, and anyone who doesn't write good briefings for their team won't write them for the AI either.

6. What does it need to look like to be processed in the next step?

Then comes the last thought that separates professionals from tinkerers, namely the question of what happens to the result afterwards.

If the table goes into a dashboard, it needs fixed column names. If the text goes to the client, it has to be written in your tone of voice, and if the analysis feeds into sales forecasting, the categories have to match those in the CRM.

Only once you know these handovers is it even worth thinking about automation. Before that, you're just automating chaos, and automated chaos is, in the end, just faster chaos.

How to tell you're collecting agents instead of getting work done

Here's a quick self-test, and feel free to be honest with yourself, because nobody is reading along:

  • You spend more time setting up your agents than on the tasks they're supposed to get done.
  • You can't say in one sentence what problem agent number seven actually solves.
  • You only skim the results, because by now there are simply too many of them.
  • You feel productive, even though your most important task hasn't been finished for two weeks.
  • You spend longer explaining your setup to colleagues than you actually use it.

If you nodded at two or more of these points, you're in good company, because I felt exactly the same at the start; the new toy is simply too tempting. The way out isn't to switch everything off, but to take a step back and answer the questions above. After that, you're usually left with two or three applications that really make a difference and then run reliably, too.

Start with that, not with licences

Better to start with that than to buy n8n or Make licences.

I see it time and again in workshops: a team has bought three automation tools, two AI subscriptions and an agent framework from a YouTube tutorial. Yet nothing has been built that actually runs in day-to-day work, because nobody clarified beforehand which workflow is actually the biggest problem.

That's why tools are always the last step and never the first, and the order looks the same every time:

  1. Pick a task that really hurts
  2. Break the workflow down into individual steps
  3. Decide which steps AI takes over and which you do yourself
  4. Define the output format and the handovers properly
  5. Test everything manually with a simple chat
  6. Only once that works, start automating

The fifth step is the most underestimated of them all. If a workflow doesn't even work while you play it through by hand in a chat window, it won't get any better with ten agents, just more expensive.

In a free video with a follow-along PowerPoint, I'll show you how to work through the first steps in half an hour and apply them to your own day-to-day work.

An army of agents for goods receiving?

And no, you simply don't need an army of agents to cross-check goods received at the rolling mill.

I'm choosing this example very deliberately, because it's so wonderfully unsexy that no influencer will ever make a Reel about it. And yet it's exactly the kind of task where AI saves mid-sized companies real money.

The delivery note comes in, the order is in the system, and someone compares line items, quantities, batches and dimensions. If there are discrepancies, a query goes out to the supplier. So it's a clear, repeatable workflow with an unambiguous result: it either matches or it doesn't.

All you need for this is a cleanly built process with a model that reads the documents, a rule that does the matching, and a person who takes a look when there are discrepancies.

That has nothing to do with an agent orchestra; it's simply a tool that does one single thing well. If that one thing saves two hours of work every day, it's worth more than any demo with 27 bots on a stage in San Francisco.

These are exactly the kind of workflows your team builds in the AI Hackathon: two days, a real process from your company, and at the end a working prototype rather than an agent demo.

Fewer agents, more clarity

I certainly don't want to talk you out of experimenting with agents, so go ahead and try out what works. It's an exciting technology, and some teams will genuinely use it to build things that were impossible just two years ago.

But don't let anyone tell you that you can't play along anymore without an army of agents. Most of the companies I work with don't have too little AI, but too little clarity about what they actually want to use it for.

So the crucial question isn't how many agents you have running. Rather, it's which one task would make your working life noticeably easier if it were reliably taken care of from next week on.

If you have an answer to that, you're already further ahead than most screenshots with twelve terminal windows.

If you can't prioritise, you don't need more agents. You need a plan first.

Which task is the frog you've been pushing ahead of you for weeks?

Back to overview
Pfeil nach oben