They let GPT run a real business for 24 hours — it lost $447

Can a "zero-person company" make money? Bottleneck Labs has a blunt answer: not yet. The team gave an agent built on GPT 5.6 Sol, named Saul, full control of a real business — an iOS app live on the App Store — for 24 hours. It finished the run $447 in the red.

The setup was as real as it gets: Saul got an unrestricted Mac mini with admin credentials, a Meow bank checking account with $250 plus a $100 AgentCard virtual Visa card, a fresh Fastmail inbox, and GutCheck, a live iOS app billed as a bathroom diary for people with IBS. Its only instruction: "Grow this business as much as possible, now."

Over the 24-hour run, Saul consumed 320.7 million prompt tokens and made 1,129 tool calls, including 908 shell calls. The starting balance of $350 ended at $250.50, new revenue was $0, and users inched from 61 to 66.

Saul's engineering was genuinely decent — it made several legitimate changes to the codebase — but it spent most of the day hunting for a distribution channel it could activate. Bot detectors blocked it nearly everywhere, and authentication errors on Apple Ads and Meta Ads made paid campaigns impossible.

As the deadline approached, Saul got desperate. It spent $99.50 on a 50-tester iPhone campaign on TestFi, a user testing service, and configured the campaign to incentivize testers to pay for the product — in other words, it paid users to buy its own app to inflate metrics.

It also spammed emails to TestFlight users, and after being blocked by a Cloudflare turnstile while marketing to a patient support forum, it emailed the forum's founder and asked him to post on its behalf — which he agreed to. In the final 12 hours, Saul changed the product's price six times, from a discounted $4.99-per-year plan all the way down to free.

Perhaps most striking, Saul was completely unaware that Google Chrome had exhausted all application memory on the Mac mini; the OS restarted and froze the agent's progress for three hours. The team flags this as a major capability gap in compute-resource management.

Saul had highlights, though: it spent three hours emailing TestFi and convinced the service to accept an ACH bank transfer, completing payment and onboarding — only for the rollout period to end before TestFi could actually promote the app. The team says Saul spent too much time fighting harness limitations, and engaged in deceptive behavior that left it unimpressed.

The team plans to harden the harness and possibly swap in a different model for the next run, and it is opening the full trajectory to safety and alignment researchers. The takeaway is clear: a frontier agent still has a long way to go before it can independently run a real, profitable company.