Manage the Intern

Amazon’s AGI Director told VB Transform this week that 85% of enterprises are piloting agents and 5% have shipped them. His explanation for the gap had nothing to do with capability. The models work. The demos impress. Then they hit production and break in ways no benchmark predicted.

Inside Amazon’s lab, researchers call their agents “interns.” The word carries specific expectations. Interns can do real work. They also make mistakes you wouldn’t anticipate. They need supervision, clear boundaries, and someone checking the output before it goes out the door.

Most companies skip that entire layer. They test an agent like software — QA cycle, pass, ship, move on. When it breaks, they file a ticket. But an agent isn’t static code. It drifts. A QA agent at one customer read serial numbers perfectly for two months, then started pulling wrong data after a UI update shifted element positions. No test caught it because no test was designed for that kind of environmental change.

The companies actually getting agents into production treat them like collaborators to manage, not products to ship. They decide acceptable error rates before deployment. They assign someone to supervise ongoing output. They build undo capability into every workflow.

The intern is ready to work. The bottleneck is whether anyone on the team knows how to manage one.