Give the Agent a Job Site

OpenAI’s hosted sandboxes are easy to mistake for developer plumbing.

They are more than that.

They are a sign that agent work is becoming physical in the operational sense. Not physical like a desk. Physical like a place where files live, code runs, tools connect, costs accrue, and someone has to clean up the mess after.

That is the part leaders should pay attention to.

A chatbot can get away with being vague. An agent with a workspace cannot. Once it can touch files, install packages, create artifacts, run jobs, and hold context across a session, the question changes.

Not: what prompt should we use?

Where does it work?

What can it touch?

What does it leave behind?

Who reviews the output?

Who pays for the run when it loops for an hour?

Most companies are still trying to buy AI like software. Pick a tool. Assign seats. Announce adoption. Hope productivity appears.

Agents punish that laziness faster.

If the work has no defined site, the agent will improvise one: a folder, a thread, a shared drive, a temporary container, a spreadsheet nobody owns. That may be fine for a demo. It is dangerous as a habit.

So start smaller and more boring.

Give one agent one job site. Define the inputs, the outputs, the permissions, the cost boundary, and the review path.

Then let it work.

The prompt is not the foundation.

The place is.