A junior engineer who does not understand the task asks a question. An agent that does not understand the task produces confident output.

That single asymmetry undoes most of what management is for. Nearly every technique a good manager has — creating enough safety that people admit confusion, reading the room for the thing nobody will say, developing someone over a year — exists because human uncertainty is quiet and has to be drawn out. Agent uncertainty is not quiet. It is invisible, because it presents as an answer.

So when people say agents are like junior engineers, they have the analogy backwards in the one way that matters. A junior gets better and tells you when they are lost. An agent does neither, unless the system around it makes lostness impossible to hide.

The claim

The leadership layer over agentic work is thin. Thinner than anyone selling “AI transformation leadership” will admit, and thinner than it was over the same work done by people.

I want to be precise, because thin is not absent. Two things are collapsing and one is not:

  • Coordination collapses. Most of what fills a delivery manager’s week is moving information between people who cannot see each other’s work. Agents do not need that.
  • Development collapses. You do not grow an agent. You replace it, constrain it, or change what it is permitted to do.
  • Accountability does not move at all. Someone still signs. The question who approved this, on what evidence has exactly the same answer it had in 1995.

What is left after the first two collapse is the third — and the third was always the hardest part of the job, just the least visible.

Why this is not the usual argument

The common version of this claim is a headcount argument: agents mean fewer people, so fewer managers. That is a cost story and I find it uninteresting, partly because it is probably wrong in the medium term and mostly because it misses what changed.

The interesting change is that the failure modes inverted.

Managing people is largely a problem of surfacing. The information you need is in someone’s head and social cost keeps it there. Almost every management practice worth the name — one-to-ones, blameless postmortems, psychological safety — is a technique for making it cheaper to say a true thing.

Managing agents is the opposite problem. Nothing is withheld, and nothing is hidden by social cost. The difficulty is that the output arrives fluent, complete and undifferentiated by confidence. There is no tell. A wrong answer has the same texture as a right one.

So the technique inverts too. You do not build safety to make truth cheaper to say. You build structure that makes an unverified claim impossible to act on. Not a culture. A boundary.

That is a different job. It draws on some of the same judgment and almost none of the same skills.

Leading without line authority

Long before any of this, I spent a stretch as the external party inside someone else’s organisation, working across three delivery teams inside a programme of more than ten. About thirty-six people, and no line authority over any of them.

The programme had moved to a functional language none of us had used, mine included. Most of the team picked up the syntax quickly and then stalled where it counted: connecting it into working software. And the work ran across verticals, so our output waited on another domain’s input and theirs on ours.

I could not direct anyone. What I had was a method — test cases before implementation, and the test data agreed across teams before the dependency landed rather than after. Sprints stopped ending in a queue.

It spread because it worked, not because I could tell anyone to use it.

At the time I thought of that as the constrained version of leadership: what you do when you lack the authority to do it properly. I now think that was exactly backwards. Leading by method rather than mandate is not a workaround for missing authority. It is the only kind of leadership that survives contact with a system where most of the work is done by something you cannot instruct, motivate, or develop.

Method scales to agents. Mandate does not. An agent has no career, no morale and no loyalty to trade on — the entire currency of positional authority is worthless. What remains is whether the standard you set is good enough that work built to it holds up.

What the role still requires

If coordination and development collapse and accountability does not, the job reduces to three things. They are unglamorous and they are the whole role.

Define the boundary before the work, not after the output. What an agent may propose and what it is permitted to execute are different questions and must be answered separately, in advance. Approval attaches to an action, not to an actor — otherwise “approve this” quietly becomes “approve this agent”, and the boundary widens by use rather than by decision.

Decide what counts as evidence. This is the part that most resembles the old job and the part most people skip. Somebody has to say what proof looks like before anyone needs it, and has to be willing to reject work that lacks it. Every system I have built that stayed honest had this. Every one that drifted did not.

Stay the person who signs. The accountability does not distribute across the system no matter how good the system is. If you cannot explain a decision your agents made, you have not delegated it — you have abandoned it.

Notice that none of the three is about the agents. They are about the standard the work is held to, which is why the layer is thin: one person can hold a standard across a great deal of work. One person cannot coordinate it.

What this does not mean

It does not mean people management is obsolete. It means it is a different job that happens to sit nearby, and hiring for one while describing the other produces someone who does neither. Humans still need growing, and the people doing this work need it more than most, because the failure modes are new and nobody is fluent in them yet.

It does not mean thin is easy. A thin layer is load-bearing. Fewer people between the standard and the work means fewer places for a bad standard to be caught before it propagates. The blast radius of poor judgment increases as the layer thins — which is an argument for judgment, not against thinning.

And it does not mean the layer disappears. Somebody defines the boundary. Somebody decides what evidence means. Somebody signs. Those three do not automate, and I have not seen a serious argument that they will.

The uncomfortable part

The honest consequence of all this is that a great deal of what currently gets called engineering leadership is coordination work that is about to stop being needed, described in the vocabulary of judgment work that is about to become scarce.

Those two are not the same, and the market has not yet learned to tell them apart. It will, and the tell will be simple: can this person say what they would refuse to ship, and why, before they have seen it?

That question was always the job. It just used to be possible to have a career without ever being asked it.

A working example of the second point — deciding what counts as evidence, and building a system that refuses to claim otherwise — is in Two proofs, or it does not exist.