A junior engineer who does not understand a task can ask a question. An agent can misunderstand the same task and still give you a confident answer.
That difference has changed how I think about leading work done with agents. With people, much of the work is creating enough trust and context for uncertainty to surface. With agents, the uncertainty can arrive looking like a finished answer.
I do not think of agents as junior engineers. A person develops judgment, learns through experience and has something at stake in the work. An agent can respond to better context or feedback within a task, but it has no career to develop, morale to rebuild or responsibility for the consequences.
A thin layer, with real responsibility
My view is that the leadership layer over agentic work is thin, but load-bearing. I am talking about the work of directing agents, not the work of leading the people around them.
Some coordination becomes much cheaper. Agents can exchange structured context without a meeting, provided the system makes that context available and keeps it reliable. The work of developing a person also has no direct equivalent: I can change an agent’s model, tools, context or permissions, but I am not helping it grow into a role.
Accountability remains with people. Someone still decides what should be done, what authority to grant and what evidence is enough to accept the result. That responsibility does not shrink with the amount of coordination.
I see this as a change in the kind of leadership required, rather than a headcount argument. The scarce part is judgment: knowing which work should proceed, which boundary must hold and what I would refuse to ship.
Why the management analogy breaks
With people, important information can stay unspoken. Someone may be unsure, disagree with a decision or see a problem they do not feel able to raise. Trust, feedback and psychological safety help make those things discussable.
An agent’s confident answer presents a different problem. Fluency gives me very little basis for deciding whether the work is right. Encouragement will not fix that. I need a process that makes uncertainty visible and prevents an unverified claim from becoming permission to act.
The distinction matters when defining the role. Leading people and directing agents draw on some of the same judgment, but they are different disciplines. Treating them as interchangeable leaves important work unattended.
What I learned without line authority
Earlier in my career, I led one team in a programme of ten delivery teams. I could direct my own team; the others did not report to me. Much of my work was keeping the teams and management aligned.
The programme had moved to a functional language none of us had used. Learning the syntax was only the beginning. The harder part was connecting the work into functioning software while dependencies crossed several teams.
We wrote test cases before implementation and agreed test data across teams before the dependencies arrived. That gave people something concrete to work towards. Sprints stopped ending in a queue of integration work, and the method spread because it helped.
At the time, I thought of this as leadership under constraint: what I could do without authority over everyone involved. I now see the method itself as the useful contribution. People could understand it, use it and judge whether it worked.
That lesson carries into agentic work. Position has no value to an agent. The objective, context, tools, permissions and acceptance criteria are what shape its work.
What I still have to decide
Three responsibilities remain with me.
Define the boundary before work starts. What an agent may propose and what it may execute are separate questions. Approval should attach to an action and its scope. It should not quietly become trust in the agent to perform similar actions later.
Decide what counts as evidence. Someone has to define “done” before the output arrives. A convincing answer can feel checked when it has only been produced. For important work, I want the evidence clear enough that I can reject a result that lacks it.
Remain accountable for the decision. If an agent takes a consequential action, I need to be able to explain why it was permitted, what supported it and what happened afterwards. The agent can help create that record. It cannot take responsibility for it.
A small leadership layer can hold those standards across a great deal of execution. Its size says nothing about how difficult the judgment is.
People still need leadership
None of this removes the need for people management. People need trust, development, feedback and room to challenge a premise. Teams building agentic systems need that support while they learn unfamiliar failure modes.
Nor does a thin layer make the work easy. Fewer people between a decision and its execution means fewer opportunities to catch a bad standard before it spreads. The consequences of poor judgment can grow even as coordination becomes simpler.
That is why I separate coordination, direction and leadership. Agents can reduce the first and execute against the second. People remain responsible for deciding what is worth doing and what they are prepared to stand behind.
The question I would ask someone leading this work is specific: can you tell me what you would refuse to ship, and why, before you have seen the output?
A working example of defining evidence and holding that standard is in One proof isn’t enough.