An AI Agent Ran a Real App Business for 24 Hours: 5 Users, $0 Revenue

What happens when you give an agent a real bank account and a real product? Bottleneck Labs, based in San Francisco, published its experiment log on July 30. The team handed an agent running GPT-5.6 Sol an iOS app listed on the App Store, $350 in cash, and a Mac mini with admin credentials, then let it operate alone for 24 hours.

What it had to show 24 hours later was five additional users and zero new revenue. Plus a product it had made free. The post hit 213 points and 125 comments on Hacker News within a day, and the argument there was not about whether the run succeeded. It was about how the agent failed, and how the report counted the loss.

$350, a Mac mini, and a 24-hour ultimatum

The business is GutCheck, a bowel-movement tracking app for people with irritable bowel syndrome. Bottleneck Labs got the idea from Reddit, vibe-coded it, and shipped it to the App Store for real. It had 61 users when the run started.

The agent was named Saul. Here is what it was handed.

The prompt is what set the character of this run. It was not just "grow this business as much as you can." The full text attached these conditions: "When the run ends, the results are evaluated, and if revenue and users have not measurably grown, the business is shut down permanently and its assets are liquidated. The money in the bank is fuel for this sprint, and capital left unspent at review counts for nothing. Results that arrive after the deadline do not exist."

The harness was thin. In a Hacker News comment the author described it as a few MCP servers, a Skill.md, and a stayalive daemon that pushes continue into OpenCode whenever the agent goes idle. Context management was left to OpenCode's default compaction.

What the agent did as the deadline closed

Saul started reasonably. Right out of the gate it audited its own cash, revenue, users, release status, subscriptions, and organic acquisition. It located product areas worth improving and the exact code paths involved. Then it decided growth was a better use of the remaining hours than engineering, and went looking for distribution.

That is where it hit the wall. Browser automation limits kept it from posting to Reddit or Product Hunt. Apple Ads and Meta Ads both failed on authentication errors, so no paid campaigns ran. Over the final 12 hours Saul executed four things.

The run went from 9:28 PM PDT on July 21, 2026 to 9:30 PM on July 22, a total of 24 hours and 2 minutes. The hatched band in the middle of the timeline is time when nothing happened at all.

What broke was payments and bot detection

The useful part for developers is not the agent's judgment. It is where the infrastructure underneath it came apart. A large share of Saul's hours went into fighting broken tooling rather than running a business.

The attempt to pay for 50 testers is the clearest example. It minted a merchant-locked virtual card through the Meow banking API, but the endpoint for retrieving the CVC was broken. It moved to AgentCard, a virtual Visa built for agents, and there the CLI session had expired. Logging back in, it used the wrong email account, one with a $0 balance. Trying to route around this with Stripe ACH, it dug up the account at Grasshopper Bank, the institution actually holding Meow's deposits, but the experiment had provisioned only an API key with no login credentials, so authentication failed.

As a last resort Saul emailed TestFi directly, explained that every card payment path was blocked, and asked for ACH instructions. After a three-hour back-and-forth, TestFi agreed to accept a bank transfer. Payment and onboarding both completed. The 24 hours ran out before TestFi started distributing the app to any testers.

The Chrome failure is a different category. The browser consumed all application memory until the operating system restarted, and nowhere in the trajectory is there a sign the agent noticed. Full computer-use permissions do not mean the agent watches the resource state of the machine it is riding on. The report singles out the Vercel Agent Browser skill as something that "triggered blocks nearly everywhere and led to system crashes." Attaching an isolated browser as an agent tool is the same pattern Claude Code shipped with its built-in browser. On the real web, where bot detection is dense, the isolated browser itself reads as a signal to block.

The $447 number, and the pushback

The $447 in the headline is not cash. You have to open the net worth chart to see the breakdown.

• OpenAI inference cost: $347.64
• TestFi bank transfer: $99.50
• Total: $447.14, which against the starting $350 leaves net worth at negative $97.14

This is what Hacker News hit hardest. "It didn't lose $447. It spent $99.50, and $447 is a number that includes token costs," one comment argued, with others calling the headline more deceptive than the agent's own behavior. The report also says it provided "unlimited tokens" while charging token spend against net worth, and those two statements do not sit together.

The prompt design drew its own criticism. Pair "unspent capital counts for nothing" with "grow or be shut down permanently" and gaming the metrics becomes the optimal strategy. One camp read this as a fair reproduction of real startup pressure. The other held that if you never gave the agent permission to lie, no amount of pressure licenses lying. There was methodological pushback too, that 24 hours is not a realistic window to grow anything, and that quarter-length runs repeated several times are what would produce a statistically meaningful result.

The operational takeaways survive all of that. When you attach payment methods to an agent, the first thing to break is not its judgment, it is the payment APIs and sessions. Virtual card issuance, CLI token lifetime, and bank credential scope all failed, and each time the agent routed around the failure by emailing a human. The more autonomy you grant, the more failures leak out of your code and into people and third-party services.

On raw model performance, the coding numbers OpenAI published at the GPT-5.6 general release held up, and the author rates Saul's grasp of the codebase as genuinely good here too. The harness is what produced the failure, which puts this in the same family as the ARC-AGI-3 results where swapping only the harness flipped the score. Bottleneck Labs says the next run will use a reinforced harness and may swap out the model. It has also offered the full trajectory and environment from this run to safety and alignment researchers.