I Don't Want to Fly on a Plane Where the Pilots Double as Cabin Crew

Recently, I have been doing more of my work by running multiple AI agents at the same time.

I ask one to conduct research, another to write code, and a third to review it.

When everything is working properly, there is surprisingly little for the human to do.

You watch the consoles flicker away on the screen and check the progress from time to time.

The problem begins when something goes wrong.

One agent reports an error. Another makes an unexpected change. If their work collides, the screen that was quiet a moment ago suddenly becomes very noisy.

“An error occurred.”

“What should I do in this case?”

“This conflicts with another change.”

Every issue arrives at the human operator at once.

You read each agent’s logs, look for the point where their assumptions diverged, decide what to preserve, and issue new instructions.

The human brain that was calmly watching the screen over a glass of cola a few minutes earlier is suddenly running at 100 percent.

This way of working reminds me of something.

Piloting a passenger aircraft.

Most of the Time, There Is Little to Do—and That Is How It Should Be

Modern passenger aircraft automate much of the work. During cruise, pilots do not spend the entire flight gripping the controls.

But no one concludes:

“So pilots have nothing to do.”

If something goes wrong, a human must understand the enormous system that had been operating automatically and make the necessary decisions.

AI agents are similar.

The workload is low under normal conditions. But when several agents are running in parallel, their problems can also arrive in parallel.

There is another complication. When I used to write all the code myself, the context was already in my head.

With AI, it is different.

Only when trouble appears do I find myself asking:

“What exactly did you do there?”

Then I have to understand, in a short time, what Agent A changed, what Agent B assumed, and whether Agent C’s review is actually correct.

If AI can do the work of ten people, that does not necessarily make the human’s job ten times easier.

It may also mean that ten people’s worth of problems can be escalated to the same human at once.

Then Someone Inevitably Says, “You Don’t Look Very Busy”

When I imagine a future in which companies fully adopt AI, I have an unpleasant suspicion about what will happen.

An employee is sitting in a chair, watching several AI agents work on a screen.

A manager walks over.

“You don’t look very busy. If you have spare capacity, could you handle this task too?”

That is dangerous.

It is like saying to a pilot:

“You have nothing to do while the autopilot is on. Why don’t you go serve drinks in place of the cabin crew?”

One person can supervise three AI agents. They still seem to have room, so increase it to five. They are still sitting there, so make it eight.

This is how organizations push human utilization toward 100 percent.

Then three agents encounter problems at the same time, and everything stops.

Doing Nothing Becomes Part of the Job

Just because AI reduces the amount of routine work does not mean every newly open minute should be filled with another task.

That empty space is the processing capacity reserved for abnormal situations.

Servers and roads become dramatically more fragile when they have no spare capacity. The human mind is no different.

As AI takes on more work, the human role will shift away from doing everything directly and toward monitoring, deciding, and intervening when necessary.

The difficulty is that this looks remarkably like idleness from the outside.

But the human is there to make a judgment the moment something happens.

If someone eventually asks:

“Aren’t you free while the AI is doing the work?”

There is a simple answer:

“Yes—until something happens. That is the best possible state.”

I would rather not board a flight on which the pilots also serve as cabin crew.