Your Agents Have No Manager

He had five agents.

None had a manager.

By Friday, the whole setup felt like a night shift with no supervisor.

One draft cited dead links.

Another solved the wrong problem.

A third kept coming back with questions because nobody had taught it what done looked like.

From the outside, it looked like leverage.

From the inside, it was chaos with better branding.

He thought he needed better prompts. A stronger model. One more tool.

What he actually needed was the job nobody wants when AI enters the room: management.

Everyone Wants the Output

Harvard Business Review framed the seduction perfectly. A vendor shows a seamless demo. The agent triages support tickets, updates records, drafts a proposal, routes it for approval. The room asks the obvious question: how fast can we roll this out?

Of course they do.

The promise is intoxicating. More output without more payroll. More coverage without more meetings. A tiny company that suddenly feels staffed.

You do not want more people to manage. You want the busywork handled. You want your time back. You want the company to feel bigger without getting heavier. That is why the promise lands.

But the deeper fantasy is escape. Not just escape from repetitive tasks. Escape from the messy human job of training, reviewing, correcting, and holding a standard.

Here is the problem. If you hired a junior employee, gave them no job description, no definition of good work, no review loop, and no escalation path, you would not call the result disappointing. You would call it negligent.

Yet people do exactly that with agents, then act shocked when the work comes back strange.

The model is not rebelling. It is revealing how vague the system already was.

The Real Bottleneck Is Management

This is not future talk. Microsoft's Work Trend Index says 75% of global knowledge workers already use AI at work. Even more revealing, 78% of AI users are bringing their own tools to work, and 52% say they are reluctant to admit using AI on their most important tasks.

That last number should make you sit up.

It means a huge amount of AI labor is already happening in the dark. Not assigned clearly. Not reviewed systematically. Not folded into a real operating model.

Google is already living the next phase. According to Business Insider, the company's internal tool Agent Smith became so popular that access had to be restricted as employees piled in.

The agents are here. The management layer is the missing piece.

Most people think agent problems are model problems. Sometimes they are. More often they are supervision problems with glossy screenshots.

When the output is inconsistent, builders reach for a new model. When the real move is usually older and less exciting: narrow the role, define the handoff, set the standard, check the work.

More Agents Is Usually a Confession

Anthropic's guide to building effective agents says something refreshingly unsexy: the most successful teams were not using complicated frameworks. They were using simple, composable patterns, and they started with the simplest solution that could work.

That matters because the modern temptation is to build a swarm before you can manage a single seat.

One agent researches. One writes. One summarizes. One posts. One scores. One reviews. It looks sophisticated. It also gives you six different places to hide from one hard question: what, exactly, is the job?

A bad manager keeps hiring before the first person has a real role.

Builders do the same thing with agents.

Every extra agent can become a delay tactic. Instead of deciding how the work should be judged, you add another layer and hope the stack figures it out for you.

It will not.

A swarm is not a system.

It is often just indecision arranged into boxes.

Anthropic makes a second distinction most people skip. Workflows are better for predictable, well-defined tasks. Agents make sense when flexible, model-driven choices actually matter. If you cannot explain the path, the checkpoints, or the acceptable output, you probably do not need more autonomy. You need better process.

Unmanaged AI Becomes Shadow Labor

This is where the seduction turns.

People love the idea of AI employees because AI feels easier to control than people. No awkward feedback. No salary conversation. No bruised ego in a one-on-one. You type. It responds. Clean. Efficient. Frictionless.

Until it is not.

Because the moment an agent touches real work, the old management questions come back with a new face. What is this role responsible for? What counts as acceptable? When should it stop and ask for help? Who owns the mistake if it does not?

If nobody can answer those questions, the agent is not saving the team from management. It is creating shadow labor.

Shadow labor is work happening inside the business without clear ownership, clear standards, or clear disclosure.

That is why the Microsoft numbers matter so much. When people are reluctant to admit AI is touching important work, you do not have leverage. You have invisible production with weak accountability.

And invisible production has a habit of becoming visible only when something breaks.

Seat. Score. Review. Escape.

If you want agents to help, every agent needs four things before it gets a real job.

A seat. One narrow responsibility. Not “help with marketing.” Something a sane person could judge in under a minute.

A score. A clear definition of good. Accuracy rate. Acceptance rate. Tone match. Time saved. Pick something that makes bad output obvious.

A review. A human checkpoint at a specific stage. Not vibes. Not “I'll glance at it later.” A real review point.

An escape hatch. A rule for when the work comes back to a human. Uncertainty, edge cases, factual risk, legal risk, brand risk. Define the line before the miss.

That is management.

Not glamorous. Not viral. Not magical.

But it is the difference between leverage and drift.

Once those four things exist, the agent can be useful. Before they exist, the agent is just amplifying the ambiguity already in the room.

Start With One Seat, Not a Swarm

Most businesses would get more from one well-managed agent in one boring seat than from six semi-autonomous agents roaming around the org chart.

Start where the work repeats, where the quality bar is legible, and where failure is cheap enough to study.

Maybe that is first-pass support triage. Maybe it is CRM cleanup. Maybe it is turning call transcripts into follow-up drafts. Maybe it is pulling raw research into a format a human can quickly approve or kill.

Run one seat until the output gets boring.

Boring is the goal.

Boring means the role is clear. The standard is clear. The handoff is clear. The review is light because the management was heavy up front.

Only then add another seat.

The point is not to build a little robot company for the screenshot.

The point is to remove low-judgment work so your judgment can go where it actually pays.

The transformed operator does not wake up to twelve mysteries. They wake up to one queue, one review step, and one set of misses worth studying. The system stops feeling haunted. It starts feeling run.

That is the real relief in all this.

If your agent stack feels messy, it does not automatically mean you need better tools. It may mean you finally found the real bottleneck.

Not prompting.

Not model selection.

Management.

Build the seat first.

Then hire the agent.