The agency ladder
When do you let the agent press publish?
Dorka Nagy-Józsa, Horizontal Thinking (2027)
Horizontal Thinking calls this the central leadership question of the next five years. Not “can the agent write the report?” - it can. The question is at what point the thing the agent made goes out, to the client, to production, to the regulator, without a human having read it first. Right now almost everyone answers it by accident.
Five rungs
Every task an agent does for you sits somewhere on the agency ladder. Draw it for your own workflows, on paper, because vague trust is exactly the thing the ladder is there to kill.
- DraftThe agent produces; you rewrite. An intern with infinite stamina. This is where almost the entire market is parked, including people who describe themselves as “AI-first.”
- ReviewThe agent produces; you read the whole thing and approve or edit. You're an editor now - the ratio already flipped: it does, you check.
- Veto windowThe agent produces and queues for action; it ships unless you stop it within some window. You sample, the way you sign expense reports: you read three, trust the rest, and everybody knows you might read theirs.
- Publish and reportThe agent acts, then tells you what it did. You read summaries, audit occasionally and manage by exception - how you treat a competent senior team member, if you're honest.
- AutonomousThe agent acts within a mandate and doesn't report each act. You review outcomes at some cadence, the way a board reviews a CEO.
Two rules
No rung is “correct.” A tweet and a wire transfer do not belong on the same rung, ever.
The placement is a leadership decision, and it expires. Capabilities move monthly; a placement you made in January is folklore by June. The leaders who will look brilliant in 2030 are the ones re-running the placement decision on a schedule - not the ones who “trust AI” or “don't trust AI” as a personality trait. Default expiry: 90 days.
One client - a CEO of 800 people, very proud of the “AI-first” slide in his town hall deck - placed his tasks on the ladder. Every one of them was parked at the bottom rung. He'd bought a Formula 1 car and was using it to heat the garage.
Ownership does not climb the ladder
Every rung answers one question: whose name absorbs what happens next? You can delegate the button at any rung you like. The name never goes with it. If the agent publishes libel at rung 4, the court and the client will be strangely uninterested in your system architecture.
You didn't delegate ownership; you just changed how late you find out about it.
“The agent did it” is - legally, commercially, morally - the 2030 version of “the intern did it.” Which is why the book has little patience for one fashionable sentence: “We have to keep the human in the loop.” The human belongs on the name, not in the loop.
How to calibrate trust
Start lower than your ego wants. Pick tasks where errors are cheap and reversible - internal summaries, first-draft responses, data pulls - and consciously push them up one rung past your comfort. The output doesn't need the promotion. Your calibration does.
Expect your best instincts to fight you. A manager who spent twenty years developing sharp judgment does not automatically defer to a system that can't explain its own reasoning. Don't bulldoze the instinct; feed it data: sampling, spot-audits, error logs.
Trust that isn't built on evidence isn't trust, it's fashion.
The household benchmark: you don't read the toast before it pops. That's rung 5. It took the toaster fifty years to earn it. Your agents are applying after eighteen months.
The fire drill: a keep-human list
Every rung you climb makes something atrophy. In the best-known deskilling study so far, experienced endoscopists' unaided detection rate fell six percentage points within months of routine AI assistance. Let the judgment-adjacent tasks atrophy completely and you haven't automated your work - you've automated your ability to check the automation.
So the mature version of “press publish” comes with a keep-human list:
- Pick what stays human - by name.The machine may well be better at some of these. Irrelevant: someone has to be able to tell when it is wrong, and Tuesday still has to happen when the tool is down.
- Drill them.Once a month, somebody drafts the thing by hand, models the thing by hand. It's slower. So is the fire drill.
- Log the machine's misses.A team that can describe how its agents fail has earned its rung placements. A team that can't is at rung 4 on faith.
Try this on Monday
List ten tasks an agent already does for you - or could by Friday - and place each on the ladder, in writing. Writing placements down converts vibes into policy. Promote one cheap, reversible task a rung. Demote one where you've been trusting on autopilot rather than evidence; you'll learn more from the demotion. Put the 90-day review in your calendar now. And write down, somewhere you'll find it in five years, the date of the first thing you let an agent publish, unread. There will be a date. Better yours than the one your competitor picks for you.
Before you climb, it helps to know which track you are on: leading AI agents is not leading people.
In the book
This page draws on Chapter 13 (When Do You Let the Agent Press Publish?) 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
What is the agency ladder?
A five-rung model from Dorka Nagy-Józsa's book Horizontal Thinking for deciding how much an AI agent may do without a human: 1 Draft (the agent produces, you rewrite), 2 Review (you read everything and approve), 3 Veto window (it ships unless you stop it), 4 Publish and report (it acts, then tells you), 5 Autonomous (it acts within a mandate and you review outcomes).
How much autonomy should you give an AI agent?
It depends on the task, not on how much you trust AI in general. A social post and a wire transfer never belong on the same rung. Start with cheap, reversible tasks, move them up one rung at a time, and base trust on sampling, spot-audits and an error log. Re-run every placement on a schedule; the book suggests a 90-day expiry by default.
Should you keep a human in the loop?
The book argues that the human belongs on the name, not in the loop. Whatever rung a task sits on, ownership of the outcome stays with a named person - climbing the ladder only changes how late that person finds out. Reading everything forever keeps every task at the bottom rung and wastes the tools.
What is a keep-human list?
A short list of capabilities a team deliberately does not fully automate, even where the machine is better, so that people can still tell when the machine is wrong and work can continue when the tool is down. The list is drilled by hand on a schedule, and the machine's misses are logged.
Where does the agency ladder come from?
From Chapter 13 of Horizontal Thinking: Leadership after the end of the specialist era by Dorka Nagy-Józsa, PCC, published by Y2Y Ltd. on January 1, 2027.
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.