Stop working

Hi, and happy Tuesday.

I’m increasingly finding that the most important skill in the AI era may be learning how not to work.

This is not the message most employers expected to hear after spending several million dollars on AI.

Nevertheless, here we are.

A growing number of advanced AI users are spending less time writing, researching, coding or building things themselves.

Instead, they give an AI a goal, let it work for an extended period, periodically inspect the results and intervene only when something has gone wrong.

This emerging practice is sometimes called botsitting.

The clearest examples of this new working style are commands such as /goal in the ChatGPT desktop app (in the Codex section) and /loop in Claude Code. Microsoft Copilot does not currently appear to have a direct equivalent, but we can expect one to arrive in the coming months.

Botsitting is similar to babysitting, except the baby can write software, read 300 documents and accidentally email the findings to the entire company.

Alex Karp, the CEO of Palantir, recently accused parts of Silicon Valley of “chillaxing over their latte” while enterprise customers became increasingly furious about AI products that generated large bills but little business value.

There is some irony here.

Most enterprises are not yet advanced enough to be furious that their employees are chillaxing while AI does their work.

They are still furious that the AI is not doing much work at all.

But that will change.

As AI systems become capable of working independently for longer, an increasing proportion of our jobs will involve setting work in motion and then resisting the urge to interfere with it.

Botsitting versus botsh*tting

Botsitting means supervising an AI agent while it completes a substantial piece of work.

• You define the objective.

• You provide access to the necessary information and tools.

• You establish boundaries.

Then you allow the AI to continue until the work is complete or until it encounters something that genuinely requires human judgment.

Botsh*tting is what happens when this is done badly.

• You give the AI a task.

• Thirty seconds later, you become uncomfortable with its approach and interrupt it.

• You add three new requirements and start again.

• You watch every sentence appear.

• It is still wrong, so you correct the AI over and over.

After forty minutes, both you and the AI have produced less than either of you could have achieved independently.

Botsh*tting is essentially micromanagement at machine speed.

It’s exhausting.

How botsitting works

Instead of asking the AI to complete one small step, you give it a larger objective or recurring responsibility.

The key trick to avoiding botsh*tting is defining what “done” means.

If the AI cannot distinguish success from failure, it will continue working with the confidence of a consultant whose project has already exceeded budget.

Once the objective and definition of “done” are clear, the next challenge is quality control without constant human intervention.

One technique that works particularly well is asking the AI to criticize its own work before letting you see it.

For example:

You can also have the AI adopt a distinct reviewer role, effectively separating the person producing the work from the person trying to find flaws in it.

All of this is a significant change in what productivity looks like.

The most productive employee may no longer be the person producing the most visible activity. It may be the person with six agents working quietly in the background.

From a distance, that person may appear to be doing very little - watching their screen with the keenness of a security guard. 

It may look like they are “chillaxing.” But the important distinction is that somewhere nearby, the bots are still working.

Best,

Dino