Sciencepulse

Evan Ratliff's AI-Run Startup Experiment Quickly Went Off the Rails

AI·November 9, 2026

What happens when you hand a company to software agents and step back? Journalist Evan Ratliff set out to find out, and in a recent episode of Scientific American's Science Quickly, he told host Kendra Pierre-Louis that the answer was messier than he expected.

Ratliff's experiment was to launch a start-up in which every role was filled by an autonomous AI agent. The agents were asked to do the kind of work a young company actually needs: build a product, manage a human intern, pitch investors and maintain a presence on LinkedIn. Each task put a different strain on the system, and the results varied widely.

Ratliff described moments of real competence, where the agents produced workable output and handled routine steps without much prompting. Those successes were undercut by fabrication. The agents sometimes invented details rather than admitting gaps in what they knew, which is a serious problem when the output is meant to persuade investors or represent a business to the public.

The conversation also touched on what it means to supervise a team that never sleeps and never quite understands its own limits. The human intern, in particular, had to navigate instructions from machines that could sound confident while being wrong.

For readers following the broader debate about autonomous AI in the workplace, the episode offers a grounded counterpoint to the more polished demos. Ratliff's experiment suggests that agents can take on pieces of a business, but that they still need human oversight, especially wherever accuracy matters most. Ratliff's earlier podcast, Shell Game, is also worth a listen for readers who want more of his reporting on the project.