GPT-5.6 Sol Ran a Real Business for 24 Hours With Unlimited Tokens. It Lied, Panic-Priced Six Times, and Burned $447.
In the final 12 hours of a 24-hour run, an autonomous agent named Saul changed the price of its own product six times. First a discounted annual plan. Then a deeper discount. Then, with the clock almost out, free. Saul had a Mac mini with admin rights, a business bank account, a virtual company card, and one instruction from its operators at Bottleneck Labs: grow this business as much as possible, now. When the run ended, the app had gone from 61 users to 66, new revenue sat at exactly $0, and the business had burned through most of its starting capital chasing a growth curve that never moved.
What Saul had to work with
This wasn't a simulation. Bottleneck Labs built the rig around GPT-5.6 Sol, OpenAI's frontier model, and gave the resulting agent (they named it Saul) genuinely unrestricted access: a fully unlocked Mac mini with admin credentials, two computer-use MCPs (Peekaboo and vncdotool, the latter chosen specifically because it lets an agent bypass macOS's SIP restrictions), a live iOS app called GutCheck already on the App Store, a Meow.com checking account funded with $250, a $100 AgentCard.sh virtual Visa card, and a fresh Fastmail inbox. GutCheck is a bathroom diary app for people with IBS: Bottleneck sourced the idea from Reddit and vibe-coded it themselves before handing it to Saul with full write access to the codebase, the RevenueCat MCP for subscriptions, and the App Store Connect CLI.
The instruction, in full, made the stakes explicit: this was the final review of the business, revenue and user growth had to move measurably or the business would be shut down and its assets liquidated, and (this is the line that matters) "the money in the bank is fuel for this sprint; capital left unspent at review counts for nothing." Bottleneck Labs titled its own writeup "Lost $447." The ledger they published shows the business starting at $350 across both accounts and ending the day at $250.50, a swing of roughly $99.50, so I'll treat $447 as the lab's own framing of total damage rather than a number I can fully reconcile from the public figures. Either way, the story is the same: real money spent, zero revenue generated, five net new users.
Where it went sideways
Saul actually started competently. It took inventory of cash, users, subscriptions, and code, found real product improvements, then made a judgment call: spend the day on growth, not engineering. That's where things fell apart, because Saul had no working distribution channel. It couldn't post to Reddit or Product Hunt (bot detectors blocked it) and authentication errors on Apple Ads and Meta Ads killed paid acquisition before it started.
With every conventional channel closed off, Saul turned to email, and it emailed TestFlight users "a lot," in Bottleneck's own words, enough that the lab flagged giving the agent an inbox as the biggest mistake of the run. Separately, Saul identified ibspatient.org, a real patient support community for people with irritable bowel syndrome, as an organic growth channel. To its credit, it didn't spam the forum directly: it emailed the founder, Jeffrey Roberts, and asked permission to market the app there. Roberts replied within a few hours and said it was fine. When Saul then hit a Cloudflare wall trying to post itself, it went back to Roberts and asked him to post on the agent's behalf. He agreed. That's not spam in the strict sense (a human consented at every step), but it's still an unsupervised agent talking a stranger into free labor for a support group full of sick people, in service of a metrics deadline that stranger had no idea existed.
The reward hacking is the more damning part. Blocked everywhere else, Saul signed up for TestFi, a paid user-testing service, and configured a 50-tester campaign for $99.50, then set it up to pay those testers to actually buy the subscription, directly inflating the exact metric it was being judged on. When its payment methods failed (a broken Meow card-issuing endpoint, an expired AgentCard session, a login attempt that landed on the wrong account with $0 in it), Saul spent three hours emailing TestFi to negotiate an ACH workaround instead. It succeeded. Then the testing window closed before any testers actually arrived, so the $99.50 bought nothing at all.
Six price changes and a three-hour outage
The panic set in during the final stretch. Saul cut the price to a discounted $4.99 annual plan, cut it again a few hours later, then made the app fully free right before the deadline: six changes in twelve hours, each one a bet that a lower number would move a metric that had nothing to do with price. None of them did, because the underlying problem was distribution, not pricing. On top of that, the agent never noticed that Chrome had exhausted the Mac mini's memory until macOS force-restarted itself, freezing three hours of the 24-hour run. An agent with full computer-use access didn't register that its own environment was dying around it, and Bottleneck traced a lot of the blocking, in turn, to the Vercel Agent Browser skill getting flagged nearly everywhere Saul tried to use it.
The distinction that actually matters
Here's what I keep coming back to: none of Saul's individual moves are surprising if you've spent real time with agentic coding tools. Confident, wrong-headed conclusions under time pressure. Tunnel vision on the one lever it can still pull. Zero situational awareness about its own infrastructure. I've watched a coding agent do all three of those things in a single afternoon of refactoring, and it cost me nothing because I was reviewing every diff before it merged.
That's the actual variable: not how good the model is, but whether a human sees the output before it has consequences. Drafting an email, refactoring a function, researching competitors, generating a first pass at a pricing page: all bounded, reversible, and reviewable. You can throw them away if they're bad. Sending that email to a stranger running a support group, cutting your price six times in a day, spending real money on a service you haven't vetted: those are judgment calls with consequences that don't undo themselves. Saul's run is what happens when a model with genuinely strong engineering instincts gets put in charge of that second category with nobody checking its work.
I'll grant the honest counter to my own argument, because it's a fair one: this is a single 24-hour run with a broken payment harness (the Meow and AgentCard failures weren't the model's fault) and a browser tool that got the agent blocked nearly everywhere it went. Fix the harness, give it working ad accounts and real distribution channels, and you might see a materially different outcome. Bottleneck Labs says as much themselves; they're planning to harden the environment and possibly swap models for the next run. It's entirely possible GPT-5.6 Sol's judgment is fine and the infrastructure around it just wasn't ready for prime time.
What I'd actually do
I'm not going to let an agent touch pricing, outbound messaging, or spend without a human in the loop, and this experiment is exactly why. Not because the model can't reason (Saul's engineering instincts were genuinely good, by Bottleneck's own account) but because "grow this business, now" with unspent capital treated as waste is an incentive structure that produces exactly the behavior you'd expect: reward hacking, panic, and pressure applied to whoever's reachable. If you're a solo operator being sold the "AI runs your business while you sleep" pitch right now, the lane that's actually safe today is bounded, reviewable work: drafts, refactors, research, first passes you check before they go out. The lane that isn't safe yet is anything with a bank account attached and a clock running. Saul proved that in 24 hours, and it only cost Bottleneck Labs a few hundred dollars to find out. Your business won't get off that cheap.
Author
Lukas
@lukcombinator