Managing a team of agents: leadership when roles become software

#ai #leadership #management #productivity

Part 3 of five. "AI, leadership and the human structures of work" is a series on what AI actually changes about leading people, and why those changes are choices, not inevitabilities.

An agent on your team drafts a client report. Another reviews it. A third sends it. It goes out with a number that is wrong, not obviously wrong, plausibly wrong, and the client acts on it. Monday morning, someone asks who is responsible. Point to a person. If you find yourself hesitating between the drafting agent, the reviewing agent, and the human who "just" pressed send on work three agents produced, you have found the real management problem with agents, and it is not their capability.

So far in this series: the human cost of AI is a leadership choice (piece one), and the authority in the room is no longer whoever knows the most, but whoever knows where to trust the machine (piece two). This piece is about what management becomes when the organisation stops hiring for a role and starts spinning up an agent for it instead.

Picture the team you might run in two years. Three people. Nine agents: one drafting, one reviewing, a couple monitoring, others doing work that used to carry a job title. The people wear several hats each; the agents wear the rest. Our entire apparatus for running teams, ownership, accountability, span of control, was built for humans. Agents quietly break all three.

Oversight was the job. It is not any more.

For decades, a large part of middle management was moving information and checking work: routing decisions up and down, monitoring who did what, catching mistakes before they travelled. Agents do the routing and much of the checking themselves. So the manager whose value was oversight is, honestly, in trouble, and the numbers reflect it. Middle managers grew to around 13% of the US workforce by 2022, up from roughly 9% in the early 1980s, and a good deal of that growth was the oversight work agents now absorb. The pressure is already visible: a June 2026 Harvard Business Review analysis describes middle managers being overloaded by AI adoption, and Gartner predicts that this year one in five organisations will use AI to flatten their structure, eliminating more than half of their middle-management positions. The oversight layer is not just shifting; in places it is being deleted. The research on where the role is heading is consistent: from monitoring to facilitation, from watching work happen to making it possible.

That is not a demotion. It is a harder job. Overseeing ten people is a known problem with a century of practice behind it. Enabling three people to direct a shifting fleet of agents, and staying accountable for what that fleet produces, is not a problem most managers have ever been trained for.

The three things agents break

Ownership. People take ownership of work they feel is theirs. Split a task across three people and six agents and ownership evaporates: everyone contributed, nobody owns it. The fix is not technical, it is a deliberate management act. Name a human owner for every outcome, not every task. The agents can do the work; a person still has to own the result, including the parts the agents got wrong.

Accountability. This is where the "it's just delegation" objection breaks down. Good managers already delegate, the argument goes, and org charts always adapt. But delegation, properly understood, is to a human who can be asked why, who feels the consequence, who learns and carries the responsibility next time. An agent can do none of that. It cannot be accountable. So accountability does not distribute when you deploy agents the way it does when you delegate to people. It concentrates, upward, onto the humans who directed them. The old management principle held that you can delegate authority but never responsibility. Agents make that literal and unforgiving: the more of a team's work you automate, the more exposed its remaining people become, because they now answer for output they did not personally produce. A leader who misses this will let their best people quietly absorb unbounded risk, and call it efficiency.

Reliance. Here is a result worth pinning to the wall. In a 2025 experiment, people systematically over-relied on AI advice even when their own judgment would have been better, and the driver was not laziness. It was incentives. When people were rewarded for throughput, they rubber-stamped. When the incentive was redesigned to reward good judgment about when to trust the AI, over-reliance fell. Your team will trust agents exactly as much as your incentives tell them to. Reward speed, and you will get rubber-stamping, and, per the accountability point above, you will personally own the results of it.

What managing actually becomes

Less monitoring, more designing. Who owns which outcome. What a human must still decide. How you reward judgment over throughput. And what your real span of control now is, because a manager of three people directing fifty agents has a span of attention no org chart has ever had to model. The manager's craft moves from supervising effort to engineering accountability across a team where most of the doing is done by things that cannot be held responsible.

That is a real skill, and almost nobody has it yet, which is precisely the argument for learning it early. The managers worth most in five years will be the ones who worked out, now, how to keep ownership and judgment human while the doing moved to software.

There is a limit to how far you can solve this one team at a time, though. Ownership and accountability do not only live in a manager's head. They live in how the whole organisation is structured, and most organisations are about to reach for the same structure to "manage AI." Most of them will build a bottleneck and call it a centre of excellence. That is where the series goes next.

The series: AI, leadership and the human structures of work