About forty minutes into a workshop, a woman who had barely spoken all morning raised her hand.
On the screen, a team of AI agents was building a marketing workflow. A copywriter, a researcher, a reviewer. Each with a role, each waiting for instructions.
“How do you stay in control,”
she asked,
“when so much is happening behind your back?”
It was not the question I was expecting. At that point in a session most people ask about the technology, which model, what it costs, how it connects to their tools.
She skipped all of that and went straight to the question that decides whether any of this works: how do you manage a team you cannot watch?
I did not have a good answer at the time (I mumbled something about summaries). I have been hearing versions of her question since January, and somewhere along the way I started noticing who asks it.
The room
I run workshops where people build a team of AI agents from scratch. A team, not one chatbot: a copywriter, a researcher, a marketing lead, a quality reviewer, each with a defined role, its own memory, and the ability to check the others’ work.
Everyone gets the same building blocks and the same two hours. CEOs and dog trainers, venture capitalists and tattoo artists, machine learning engineers and physiotherapists. Hundreds of people so far.
That mix is what makes the room interesting. It is close to a natural experiment: same task, same tools, and a machine that does not know who you are.
The management mirror
It took me months to see this clearly.
An agent strips away every advantage that normally tilts a workplace. In a regular organization, confidence gets rewarded, seniority gets rewarded, and a technical title opens doors before anyone checks whether the work is good. The hierarchy is always in the room, even when nobody names it.
An agent sees none of it. It responds to exactly one thing, the quality of your instructions. It does not know you are the CEO, and it does not respond to frustration, authority, or a good reputation.
It waits for an instruction it can execute.
So it works like a mirror. Whatever management skill you actually have, it reflects straight back at you. If you can break a job into steps, hand it off, and check the result, you get a working team in an hour.
If you can’t, you get a very fast, very polite pile of parts.
Nothing in between, and nowhere to hide.
I call it the management mirror.
I think it is the most honest performance review most people will ever get.
Two languages
Once you know the mirror is there, you start hearing two different languages in the same room.
One is technical talk. Which model is best, how much the tokens cost, where it runs, how to connect it to my calendar and my keys, whether this is the right tool or the other one is.
The other is managerial talk. How much should I trust it, should this agent do strategy only or also execution, does it need someone under it, why do they share one memory, where is the limit of control.
The sharpest question I have heard all year came from one participant looking at her brand new team:
“How do you verify their quality in areas where they are better than you?”
Both languages are legitimate.
Only one of them predicts who leaves with a working system.
You can hear it before anyone starts building, if you ask people what they came for. One group says “I want to build a website” or “I want to automate lead scraping.” The other group says “I want to hire a team,” “I want agents that handle the tasks I am worst at,” “I want an army of marketers,” “I want someone to do the grunt work.”
One group describes software. The other describes staff.
The second group finishes first, nearly every time.
I watched a machine learning manager, the most credentialed person in his session, spend his two hours trying to solve the workshop like an engineering problem. Halfway through he asked me, genuinely puzzled: “What’s the advantage of working this way instead of just having the AI build me the software?”
The idea that managing the tool might be a different skill from building with it had not occurred to him. He finished near the bottom. In the same session, a physiotherapist who had never opened a terminal was on her second working agent.
One man put the whole thing in a chat message, an hour into a session that was getting away from him: “In my head I’m on a rocket. In practice, all the parts are scattered on the floor.”
He was smart. What he was discovering was the distance between knowing what you want and describing it clearly enough for someone who has never seen your work to do it without you standing over them.
The part I did not plan
Now the part that makes people uncomfortable, including me.
The managerial questions come mostly from women.
I did not go looking for this. I noticed it live, and then I went back through the recorded chat logs of 80+ workshops and reread every question, looking for which language it was in.
The infrastructure questions, servers and tokens and keys and hosting, came mostly from men. The delegation and control questions, how do I lead them, how do I verify them, how do I stay in charge, came mostly from women.
Since I first wrote this down I have run dozens more workshops. I stopped counting. The pattern did not stop.
I want to be careful here, because an observation like this gets flattened the minute it leaves the room. I did not run a controlled study. Gender in the chat logs was inferred, from names and from the way Hebrew grammar gives it away.
My rooms are self-selected. And I am a man making an observation about women, which creates distortions I cannot fully see. So I am not telling you women are better at AI.
I am telling you something I think is bigger.
This is not a gender gap. It is a language gap.
The two languages are just not evenly distributed, and the reason they are not is history.
Where the language came from
For thirty years, the top of the tech org chart was the engineer, the person who could make the machine do the thing. Around that center we arranged a ring of roles, paid them less, and called their skills “soft”: project management, operations, HR, learning and development, customer experience.
These are the people who break work into steps. Who give feedback that improves instead of deflates. Who delegate, and then check, and hold a dozen moving parts in their head so the whole thing does not collapse.
Those roles are, disproportionately, where women ended up. That is who practiced the managerial language every day, for decades, while being told it did not really count.
The agents did not read the org chart.
Hadar Schwartz is a physiotherapist. She runs a Pilates business, teaches anatomy, and is doing a PhD in public health. No technical background at all.
After my workshop she froze for three days. On the fourth day she sat down and started talking to the computer, beginning with invoice scanning, because that was a process she could already describe.
In her anatomy courses she teaches movement the way she runs any complex process: what happens first, what comes next, what cannot be skipped. Describing a workflow to an agent turned out to be the same skill as describing a movement pattern to a student.
Before long she had agents scanning her inbox for invoices, answering customers on WhatsApp, running a teaching assistant for her students, and sending her a morning briefing. A role that had cost her about $1,500 a month now costs her $100.
“The most important skill I discovered, was knowing how to describe a process clearly. I’ve been teaching for years. It’s the same muscle.”
—
Shulamit Banay leads AI implementations inside organizations, and she is technical. She put it plainest:
“What helps me with AI is deep management experience. People, processes, systems. Managing an agent system is just a different implementation of the same thing.”
—
Tomer Wertheimer, who coaches people to manage their careers like a product, turned the same idea inward.
“Managing agents is mainly self-management. The agent amplifies what’s in you. There’s clarity, it produces clarity. There’s confusion, it produces confusion. This isn’t a technical skill. It’s a leadership skill.”
—
Maya Dror Melamed, a fractional marketing chief, named the history without flinching:
“Women manage many things at once. They prioritize in real time, ask questions and actually listen to the answers, give feedback and encourage. Otherwise the world would have collapsed long ago.”
The inversion
Stand back and look at the shape of it.
Once execution got cheap, judgment got expensive.
And judgment was never sitting where we thought.
The agents reward the ring and ignore the center. They turn the “soft” skill into hard currency and leave raw technical brilliance waiting for an instruction it never gets.
The people we spent a generation calling the least technical in the building were holding the key the whole time. They just did not have a team to command until now.
The most technical skill of the next decade is management.
The least technical people in the room already have it.
This is not how most organizations are responding. They are hiring more engineers, appointing chief AI officers, launching technical training, reinforcing the center while the leverage moves to the ring.
The pipeline for the most valuable skill of the AI era is running right now, entirely off the org chart, in the part of the building nobody is funding.
Which language do you speak?
If you have spent years knowing how to break a process into steps, give feedback that improves rather than wounds, and hand something off and then check it instead of hovering, and you were told those are soft skills: this is what they were for. You have been in training the whole time. Nobody called it that.
And if you are the builder, the rocket-in-your-head person: the description is the problem, and the description is learnable. And the people who can teach it to you are probably already in your room.
One thing to try this week: write down the next five questions you ask an AI, then sort them into two columns: “how do I wire it” and “how do I lead it.” Most people are surprised by which column fills up.
Reply and tell me which column won. I am collecting. And if you run rooms of your own, in any field, tell me whether you see the same split, because I want to know if it is just me.
I do not have a theory for why the pattern keeps appearing. I have the chat logs and a question I cannot put down.
If the skills most critical to managing the most powerful technology in history are the ones we underpaid and filed under “soft” for decades, what does that say about how we have been measuring competence the whole time?
I don’t know where this goes. I just know I can’t stop watching.
🛠️ Build of the Week: Tamar’s product discovery team
Tamar Shachar walked into a new VP Product role and did not wait for a research team. She built one: a researcher that takes every user interview and pulls out insights and opportunities, a competitive intelligence agent, and a strategist that helps her build the opportunity tree.
Her explanation of why it works is this whole issue in one breath (her words, from the form I sent alumni): “In product management you learn humility. A hypothesis is just a hypothesis. When you don’t need to be right, it’s easier to accept that there are things my virtual team members do much better than me.”
Managerial language, from day one. Tamar on LinkedIn.
That’s it for this week.
If this was useful, forward it to someone (real human) who was told their skills were soft.
See you next week ✌️
-- Tom
(the guy who keeps getting out-managed by physiotherapists)
P.S. This newsletter was 88% made by my ai team.
P.P.S. If you’re new here, the Loop’s back-catalog is at https://www.agentsandme.com/archive. The closest cousin of this issue is the one about hiring a team instead of one super-agent: https://www.agentsandme.com/p/jack-of-all-trades-master-of-none
P.P.P.S. I teach the full method live, two hours, the same room as everyone in this issue: https://getagents.today
P.P.P.P.S. I read every reply, the real me. Tell me which column won.