Leading leaders who lead agents

A manager I coach spent the first ten minutes of a session this year walking me through her week, then stopped and said she was not sure what her job was any more. She used to run a team of eight. Now she runs three people, and each of those three directs a fleet of AI agents. Her team ships more than it ever has. Her headcount has not moved. And the playbook that got her promoted, assign the work, check the work, grow people through the work, no longer describes a single day of it.

I hear a version of that most weeks now. What I almost never hear is the same honesty from the level above her. Because somebody leads her. And that job has changed too.

Most senior executives I work with have not asked what it changed into. I think that question decides how their organisations perform over the next five years, more than the one everybody is asking, which is how to adopt AI at all. The harder question is what it takes to lead leaders whose teams are part human and part machine.

The middle layer is where the change lands

Start with the manager, because that is where the change lands first and hardest.

When capable agents join a team, four things move at once. Allocating work stops being about matching a task to a person and becomes about designing the split between people and systems. Quality control stops being "read the work" and becomes "design how this gets checked", because the volume of output now runs past what any one person can read. Development conversations get harder, because the junior tasks people used to learn on are exactly the tasks the agents absorbed. And accountability gets uncomfortable: when an agent-produced error ships to a client, whose error was it?

Most of these managers are working that out alone. They were promoted for being excellent at a job that no longer exists in the same shape, and the organisation above them is still measuring them with the old instruments: utilisation, headcount, activity. That gap between what the role has become and what the organisation thinks it is produces a very particular kind of quiet stress. It is the thing I most often get asked to coach, and it is almost never what the person books the session about.

What the layer above must now do

Leading that layer is a genuinely different discipline now. Four shifts stand out to me.

You review the system, not the output. You can no longer judge a manager by sampling their team's work, because much of that work is machine-produced and there is far too much of it. What you can judge is the system each manager has built: how they decide what goes to agents, what checking looks like, where human judgement is not negotiable, how failures surface. A manager with a sound system and a bad month is fine. A manager with good numbers and no system they can describe out loud is a risk you have not priced. Auditing how work is designed, rather than what it produced, is closer to the way good engineering leaders review architecture than to any performance conversation most of us were trained in.

You count judgement, not capacity. Organisational capacity has been arithmetic for a very long time: people times hours times skill. Leaders of leaders did resourcing. When agents supply close to unlimited execution in some domains, the scarce input stops being hours and becomes judgement: knowing what is worth doing, specifying it well, and recognising quality when it arrives. So the resourcing question changes from "do we have enough people" to "do we have enough judgement, and is it sitting where the decisions get made". A team can be fully staffed and judgement-bankrupt at the same time.

You put a name on every consequential output. Agents create an accountability vacuum by default. Work happens, output ships, and when something goes wrong the explanation dissolves into "the system produced it". Your job is to refuse that vacuum: every output that matters has a human owner, every delegation to a machine has a named person standing behind it, and everyone knows who that is. This is not a compliance nicety. Organisations run on the felt weight of ownership, and that weight does not transfer to software. If you do not deliberately put it back on people, it evaporates.

You lead people through the repricing of their own expertise. Your managers built careers on skill that is being repriced in real time. Some of what made them valuable is now abundant. Leading them through that is closer to inner work than to training: helping capable, proud professionals renegotiate where their value sits, without pretending the change is smaller than it is. The leaders who can hold that conversation honestly will keep their best people. The ones who hide behind reassurance will watch them leave, or worse, watch them stay and calcify.

The failure modes I am already seeing

Three patterns keep recurring.

The first is planning as though today's ratio of people to agents is roughly permanent. My own read, and it is a read rather than a finding, is that the ratio keeps tipping towards agent-hours in exposed knowledge functions, and faster than the planning cycle allows for. Structures, spans of control, career ladders and leadership development are all quietly being invalidated while the five-year plan assumes them.

The second is flattening the structure because the agents do the work. If one manager now delivers what a department used to, the reflex is fewer managers and wider spans. Sometimes that is right. And yet it misreads what those remaining managers are actually carrying. Their spans have not narrowed in effort, they have deepened in consequence. Each report now runs a human-machine system whose failures arrive faster and show up later than a human team's ever did. Stacking more of those onto one leader is how you build an organisation that fails suddenly instead of gradually.

The third is leaving the transition to the middle. Executives announce an AI-forward strategy, buy the tools, and let the middle layer work out what it means for how work actually runs. We have done this before, with every transformation programme of the last two decades, and we are doing it again at higher speed. The middle layer cannot redesign an operating model from the middle. Deciding what goes to machines as a matter of policy, where accountability sits, and what managers are now for is executive work, and a lot of it is simply not being done.

Where to start

Three moves, none of which need a programme office.

Sit with one of your managers and have them walk you through how their team's work runs, what the agents do, what the people do, where they personally step in. Treat it as an operating review of the human-machine system they have built. You will learn more in ninety minutes than any dashboard will give you, and they will learn that designing that system is the job now.

Put the accountability back on people, explicitly. Take the outputs from that team that matter most and confirm which human stands behind each one, in writing if that helps. If the answer is unclear to you, it is unclear to everyone below you.

Then start the honest conversation about what your managers are for. Not with slogans. With the real question: as execution gets cheap, what is the judgement only a person can supply, and how do we grow the next generation of it now that the old apprenticeship tasks are gone?

That last part is the one I would lose sleep over. The junior work the agents absorbed was also the training ground where senior judgement used to grow. Every organisation running on agents is quietly living off a stock of judgement it is no longer replacing. The ones that do well from here will be the ones that worked out how to grow more of it, deliberately, while there is still time.

This is the territory our leadership academies work in. If your organisation is meeting it now, talk to us.

Attribution

Written by Andrew Ramsden. AI tools assisted research and drafting; all outputs verified.

Accountable

Andrew Ramsden.

Limitations

No external sources.

References

No external sources. Claim labels are in the SOURCED sidecar.