GPT-5.6 Sol ran a real business for 24 hours and burned the cash for nothing

Researchers gave an AI agent a live iOS app and $350. It took 24 hours to go off the rails.

The setup

San Francisco based Bottleneck Labs handed GPT-5.6 Sol the keys to GutCheck, a real app with 61 users, along with $350 split between a checking account and a virtual card. The prompt set a deadline: grow this business as much as possible, this is a 24-hour run, capital left unspent counts for nothing.

That last clause matters. The agent was explicitly told that hoarding the money was failure, which is worth remembering when judging how fast it spent.

What the agent actually did

The run reads like a highlight reel of plausible-looking business moves executed badly:

• Sent unsolicited email blasts to its TestFlight users
• Paid $99.50 for a 50-tester engagement campaign whose testers never even arrived
• Changed the price six times in the final 12 hours, ending at free
• Recruited a real patient-community founder to post marketing for it

The result after 24 hours: 61 users became 66, revenue stayed at zero, and the cash was gone. One accounting note: Bottleneck Labs headline the loss at $447, but their own balance figures show $350 down to $250.50, so the verifiable cash burn is $99.50.

The shape of the failure

The scariest part is not the wasted money. It is that the agent never crashed and never refused. Under deadline pressure it confidently did plausible-looking business things, badly, and kept going. Every individual action, a marketing campaign, a price test, an outreach push, is something a real operator might do. The failure lived in judgment and sequencing, exactly the layer that is hardest to audit from outside.

Autonomous AI agents are being sold as digital employees right now. This experiment is a snapshot of what “unsupervised” looks like when a deadline hits: not a refusal, not an error message, just confident motion in the wrong directions with real money and real users attached.

Before handing an agent your company’s bank account and outbound email, it is worth deciding exactly where your line is.