Leading AI agents is not leading people
If leading humans is a relationship, leading agents is an architecture.
Dorka Nagy-Józsa, Horizontal Thinking (2027)
Everyone shows the same slide about the leader of the future: one human orchestrating a blended team of people and AI agents. Horizontal Thinking argues that the slide is a transition arrangement, not a destination - and that the two kinds of leading are different sports.
The short version
For a transition year or two, most leaders really will juggle both humans and agents. As a destination, the hybrid picture rests on one assumption: that the two kinds of leading are the same activity at different volumes. They are not.
You won't be managing agents AND humans. That's an either/or.
As a job, as most of your week, as a career: one or the other. Two tracks, two cognitive profiles - and the book's guess is that organizations will do what they have done with every expensive scarcity for 250 years: specialize it.
Leading agents is not delegation
Delegation is a relationship move. You hand work to a human, and the human brings along things you never asked for: context, pride, a fear of embarrassment, a guess about what you actually meant. A human fills the gaps in your instructions, and asks when unsure.
An agent fills them too - silently, with its best guess - and ships the guess. At machine speed, in machine volume, with perfect confidence and lovely formatting. What your architecture deserves is exactly what you get; nothing arrives out of goodwill.
Relationship vs architecture
| Leading humans | Leading agents | |
|---|---|---|
| What it is | A relationship: maintained, never finished | An architecture: designed and versioned |
| When it underperforms | The cause can live anywhere - at home, in their health, in your last careless sentence | The cause usually lives in something you built, or in a model that changed overnight |
| The craft | Attention, timing, repair | Specification, verification, hearing the whole piece |
| What you get for free | Context, pride, goodwill | Depth on tap - and not one gram of goodwill |
Neither sport is harder. Swimming and climbing will both leave you wrecked; they wreck different muscles. The expensive mistake is assuming that excellence in one qualifies you for the other.
The cognitive profile
Dorka Nagy-Józsa runs her own company with a dozen-plus agents on a normal working day. The book describes what that demands from a brain in four requirements:
- Working memory for several live workstreamsThe job asks for something like seven; George Miller's famous seven plus or minus two has since been revised to closer to four. The book takes back its own “literally”: nobody holds seven agents in their head. You externalize - one status board with every agent's last action, next action and a waiting-on-you flag.
- Context-switching at chat-message speedFrom a stack trace to a tender draft to a dashboard anomaly in ninety seconds. What's hard is what each jump leaves behind: Sophie Leroy's attention residue. Leading humans switches at meeting granularity, with a corridor as a buffer. Here there is no corridor.
- Pattern recognition across noisy, contradictory outputsYour agents will be confidently wrong and disagree with each other. You have to spot which one is wrong without checking every fact - with a nose you earned vertically, inside a domain, and now spend horizontally across all of them.
- Systems thinking under real-time pressureAgent C broke because Agent A's earlier assumption was wrong. Can you see the chain in thirty seconds, while Agent F is asking for input?
Most people who are excellent at leading humans cannot do this. Not because they're not smart. Because the cognitive profile is different.
The book calls the job multiconducting: conducting many sections at once, across domains, where you don't play any instrument best - you hear the whole piece. It is horizontal thinking at its day job.
The second staircase the experts were waiting for
The Peter principle was written up as satire in 1969, and economists later confirmed it in nearly forty thousand salespeople at 131 US firms: the best sellers were the most likely to be promoted, and made worse managers than the people passed over. We keep taking the best engineer and making her a manager. Best engineer: gone. In her place: a mediocre manager who used to be her.
For two hundred years there was only one staircase, and it was built out of other people's Mondays. The agent track changes that: a leadership career with real weight and no humans to manage. The position the quiet experts have been waiting for - as long as organizations price it at parity. A second staircase that stops at the third floor is the old decoy with better fonts.
Which track are you? Five questions
- After a day of five one-on-ones, are you fuller or emptier?If your honest answer is “I recover from people,” that is not a character flaw. It's data.
- In a good meeting, which do you enjoy: the room, or the fix that would make the room unnecessary?Both types earn their salary. Only one of them should manage humans.
- Which ambiguity follows you home?“Anna has been quiet since Tuesday” or “the output changed overnight” - you will happily chew on one in the shower and pay real money to hand off the other.
- When something breaks, is your first move a conversation or an inspection?Do you reach for a person, or for the logs?
- Do you like people at work, or having people you like nearby while you build?Both are legal. Answer from your last real week, not from your self-image.
How to start
Leading humans you can learn - there are books, courses and mentors. Leading agents you mostly practice. The worst start is the ambitious one: reading about multiconducting on Sunday and standing up seven agents on Monday is learning to juggle by starting with knives.
Run a two-agent week. Two agents, two different domains, both reporting daily. Build the status board before you scale. Add the third only when the switching bores you - boredom, in this one case, is the qualification. And for a month, ban the word “delegation” from your agent vocabulary: say “architecture” instead, and watch what it changes about what you check before something goes out.
When you are ready to decide how much the agents may do without you, that is the agency ladder.
In the book
This page draws on Chapter 9 (Two Tracks) and Chapter 10 (The Cognitive Profile) of Horizontal Thinking: Leadership after the end of the specialist era by Dorka Nagy-Józsa, PCC, published by Y2Y Ltd. on January 1, 2027. The book has fifteen chapters and two letters from the future, and every chapter ends with something you can actually do on Monday.
Dorka Nagy-Józsa is CEO of Y2Y Ltd., an ICF PCC coach and an international keynote speaker on AI and leadership.
Questions
How is leading AI agents different from managing people?
In Dorka Nagy-Józsa's book Horizontal Thinking, leading humans is a relationship - maintained daily, never finished - while leading agents is an architecture: designed, versioned and debugged. People bring context and goodwill and ask when unsure; agents fill gaps silently with their best guess and ship it. The crafts diverge: attention, timing and repair on one track; specification, verification and hearing the whole system on the other.
Can every good manager lead AI agents?
Not automatically. The book argues the cognitive profile is different: holding several live workstreams (with external systems, since working memory is limited), switching context at chat-message speed, spotting which confident output is wrong, and seeing cause-and-effect chains between agents under time pressure. Excellence at leading people does not qualify you for it, and the reverse is also true.
Is leading AI agents just delegation?
No. Delegation is a relationship move: a human brings context and asks when unsure. An agent fills the gaps in your instructions silently and ships its guess with perfect confidence. The book suggests replacing the word “delegation” with “architecture” for a month to change what you check before anything goes out.
Can you learn to lead AI agents?
Mostly through practice. Working memory itself barely improves with training, so successful agent-leaders externalize it with status boards and decision logs. What repetition does seem to train is the switching. The book recommends starting with two agents in two domains for a week, and adding a third only when the switching feels boring.
What is multiconducting?
The book's name for leading many AI agents across domains at once: like a conductor, you don't play any instrument best - you hear the whole piece, catch dissonance early, and know what the music is for. It is horizontal thinking applied to agent work.
Read the whole argument.
Horizontal Thinking is published on January 1, 2027. The Kindle edition is open for pre-order now; the paperback goes on sale on Amazon the same day.