Apr 28, 2026

I'm a manager now. So are you. We're mostly bad at it.

Eighteen months of working with AI agents has convinced me of one thing: the middle managers I used to roll my eyes at were doing harder work than I knew.

By the rough count of my browser history, I have written something like four thousand prompts in the last eighteen months. I have lost mornings to small dialogues about why the agent did exactly what I asked when what I asked was not what I wanted. I have re-written the same system prompt seven times in a row. I have kicked off a multi-step workflow, walked away to make coffee, come back, and discovered the whole thing took a wrong turn three steps in because I'd forgotten to tell it the obvious thing, the thing I'd have told a junior colleague in the first thirty seconds.

What I have slowly come to realise, in a way that's mildly embarrassing to admit out loud, is that I have become a manager.

Not in the org-chart sense. I have no direct reports. HR doesn't know about my team. But I now spend a meaningful slice of my working day defining tasks for entities that are not me, providing them context, sequencing their work, reviewing their output, and getting frustrated when they get it wrong. That is the job. That is the whole job. The fact that they happen to be made of weights and not water doesn't change the structure of what I'm doing.

And I am, by my own honest assessment, mediocre at it.

This would be funnier if it weren't so widespread.

The promotion most of us dreamed of

For most of my career the cultural assumption around me was that becoming a manager was the goal. Not always articulated; just there, in the air, in the comp bands, in the org charts, in the LinkedIn updates. You started doing the work, and if you were any good, eventually you stopped doing it. You moved up. You got the chair with arms.

The survey data is more honest than the LinkedIn updates. Roughly a third of workers say they want to become managers, but when you press them, only about a quarter actually want to lead people. Most of the would-be managers want the title without the job. That is not unrelated to what comes next.

Then 2025 happened. Agentic AI gave every knowledge worker a team. Tireless, dirt cheap, faster than us, infinitely available. The promotion finally arrived for everyone, free download included.

So how is everyone doing with it?

I look around. I talk to the founders and PMs and engineers I'm friendly with. The pattern is consistent. Most people who could be delegating to agents aren't really delegating. They're using the tools as fancier autocomplete. The minority who try multi-step workflows tend to give up after a couple of frustrating sessions. Even the ones getting real value will admit, when pressed, that nobody on their team has cracked it either. The headlines are bullish. What people actually do in their day looks more like fumbling.

The promotion arrived. Hardly anyone is taking it.

The promotion paradox	Figure
Workers who want to become a manager	34%
Who actually want to lead people	23%
New managers who feel competent in their role	42%
People with natural managerial talent	10%
Managerial hires eventually judged wrong	82%
Workers who have quit specifically to escape a manager	50%

Sources: CareerBuilder workforce survey; Gallup, "State of the American Manager"; Gallup global engagement studies, 2014–2025.

Why I'm bad at this (and you probably are too)

Here is what I have learned, the hard way, about why I keep messing this up.

The agent isn't the problem. It almost never is. When I ask honestly what went wrong with a session that went sideways, the answer is some version of: I didn't think hard enough about what I actually wanted. I was vague about the output. I assumed it would know things it had no way of knowing. I gave it the headline and not the constraints. I corrected it three times instead of stepping back and rewriting the brief.

This is, I have come to realise, the same critique I have had of every bad manager I have ever worked for.

There is a whole literature on this. Mintzberg pointed out, fifty years ago, that managers spend their time on fragmented, mostly verbal work that bears no resemblance to the strategic glamour the job is sold as; Drucker, earlier, observing that the most important question in knowledge work is the one nobody bothers to ask: what is the task? Define the result. Get the brief right. Everything else flows from that.

I'd love to claim I came to this through reading. I didn't. I came to it because my agent kept giving me the wrong thing, and I kept blaming the model, and at some point I had to be honest with myself.

The asking is the bottleneck. It always has been. The literature just tells you you're not the first one to figure that out.

The team part

It gets worse, or more interesting depending on your mood, once you start running more than one agent at a time.

Now you have a planner that decomposes work, a researcher that pulls evidence, a writer, a critic. Some configuration like that. The vocabulary even cooperates with the metaphor: the literature calls it orchestration. What you're really doing is wiring up a tiny dysfunctional company. Who hands what to whom. Who has authority to escalate. What happens when the planner and the writer disagree about scope.

I have built a few of these. None of them work as well as I think they will on the whiteboard. The failures look exactly like every team I have ever been on that didn't quite work: bottlenecks at the slow specialist, conflicting outputs nobody reconciled, context that someone assumed would be carried over but wasn't. The Gantt chart of a multi-agent workflow looks suspiciously like the Gantt chart of a small startup, and the failure modes are isomorphic.

You start to appreciate, very quickly, the unsexy work of the people who do this professionally. The ones who write the actual job description. Run the standup. Notice that the brief everyone agreed on was actually two different briefs in a trench coat.

What I don't have to do

Here is where it gets darker.

Managing agents is, in almost every important respect, easier than managing people. Not slightly easier. Fundamentally easier. The agent doesn't need a holiday. Doesn't have a sick mother or a difficult divorce or a child with an ear infection on the morning of the launch. Doesn't get burned out. Doesn't quit. Doesn't sue. Doesn't post on Glassdoor. Doesn't form a union, doesn't have a midlife crisis, doesn't, on the day the merger is announced, secretly apply to the competitor down the road.

What managing agents takes off your plate:

Holidays & sick leave
Mental health & burnout
Parental leave & family events
Performance reviews & compensation conversations
Promotion politics
HR, PIPs & terminations
Interpersonal conflict
Engagement, retention & attrition
Office politics & alliances
Career-development coaching

I don't run performance reviews. I don't sit through compensation calibration. There is no HR. No PIP. No DEI workshops, no offsites, no skip-levels, no town halls, no anonymous engagement surveys with comments I can't unread. Nobody is in love with anyone they shouldn't be. Nobody is grieving.

Strip all of that out, strip out the entire emotional infrastructure that makes a workplace a workplace, and what remains is the pure, theoretical job of management. Define the task. Provide context. Set up the workflow. Review the output. Iterate.

That is it. That is the job that, allegedly, half the office has been gunning for.

And I, with all of that overhead removed, can still barely do it.

The numbers, briefly

I'll spare you the long version. The short version is that the data on actual management, done with humans in actual companies, is devastating. Forty-two percent of new managers say they understand how to succeed in their role; the other 58% are presumably faking it. Around four out of five managerial hires are eventually judged the wrong choice. Gallup's long-running studies estimate that maybe one in ten people has the natural talent for the work at all.

And the same research, just as consistently, finds that managers account for roughly 70% of the variance in their team's engagement, and that half of all employees have, at some point, quit a job specifically to escape one.

Read that again. Half. Half of the workforce has, at least once, walked out of a job because of a manager.

The manager effect	Figure
Manager's share of variance in team engagement	70%
Everything else combined	30%
Annual global productivity lost to disengagement	$438bn
Global employee engagement rate (2025)	21% (lowest in five years)
Work units in Gallup's database	2.5 million

Sources: Gallup, "State of the American Manager," 2015; Gallup, "State of the Global Workplace," 2025.

If we collectively suck this badly at the version of the job where actual humans are involved, what business do we have thinking we'll suddenly be good at it now that the humans have been removed?

The unkindest part

For most of my professional life, the cultural shorthand for an unimpressive job has been "middle management." The middle manager is a punchline: pointy-haired, weak-chinned, vaguely Dilbertish. We've all worked for one. We've all complained about them. We've all secretly believed we'd be better at it.

The polls back this up: most workers think their manager is mediocre, and most workers think they themselves would be at least good. Both things cannot be true. Someone is wrong about something.

What AI has done, quietly, is hand each of us a small private laboratory in which to test the proposition. Sit down. Open a chat window. Spin up an agent. You have direct reports now. They are infinitely patient, dirt cheap, emotionally disposable. Every excuse you have ever given for why your manager was bad has been removed from the system. There is no politics. No quarter-end pressure that wasn't yours. No HR business partner with an agenda. No fired colleague whose work you absorbed.

And mostly we, I included, turn out to be the kind of manager we always complained about. Vague when we should be specific. Inconsistent. Quick to anger when the brief we wrote produces the work we asked for. Reluctant to invest in the documentation, the system prompts, the boring scaffolding that makes the team function. We want the leverage without the discipline. We want to delegate without specifying. We want the chair without the work.

The middle manager I rolled my eyes at for most of my twenties was, it turns out, doing harder work than I knew. Probably also bad at it (most people are; the numbers are not subtle), but the part they were doing, even badly, was more than I gave them credit for at the time.

This isn't a rehabilitation of bad managers. There are real bad ones, and they cost their teams real damage. But the assumption underneath the eye-roll, that the job itself was easy and they were just gumming it up, is one I have to retract.

In the meantime, I'll be here, writing my four-thousand-and-first prompt. Specifying the output. Providing the context. Trying not to blame the agent when I forgot to tell it something obvious.