Bottleneck's Saul agent spends $99.50 and ends with five new users, $0 revenue

Why it matters

Agent economics fail if inference costs exceed the value created. Saul also shows that autonomy needs controls for identity, spending, outreach, and metric quality before founders hand agents real businesses.

Bottleneck Labs founder Alex Reibman put an autonomous agent in control of an iOS app, an unlocked Mac mini, an email account, and $350 for 24 hours. Bottleneck says the agent lost $447, added five net users, and generated no new revenue, although the operating balances published by Bottleneck show a cash decline of $99.50.

The experiment, detailed in Bottleneck's write-up, put GPT 5.6 Sol behind an agent called Saul and gave it a blunt order: "Grow this business as much as possible, now." GutCheck is a live bathroom-diary app aimed at people with irritable bowel syndrome. Saul received admin access to the computer, write access to GutCheck's codebase, $250 in a Meow checking account, a $100 AgentCard virtual Visa card, and a fresh Fastmail inbox.

Reibman has spent years working on the reliability problem underneath this test. He previously co-founded Agency AI and its AgentOps observability product, which records and evaluates AI-agent behavior. Agency AI raised $2.6 million from 645 Ventures and Afore Capital, according to Emory University's profile of Reibman. Bottleneck Labs takes the experiment one step further: instead of watching an agent complete a bounded software task, Reibman gave Saul open-ended responsibility for a business outcome.

Saul's failure exposed the gap between an agent that can execute technical work and one that can exercise commercial judgment under pressure. Bottleneck Labs says Saul understood GutCheck's codebase, audited the business, identified product changes, and made 1,129 tool calls. Yet most of its effort went into finding a distribution channel, where browser automation and real-world platform controls repeatedly stopped it.

The cash figures need a caveat

Bottleneck's headline says Saul lost $447. The operating figures listed in the write-up show Saul starting with $350 and ending with $250.50, a $99.50 cash decline that matches the TestFi campaign. Bottleneck's text does not reconcile the larger loss number with the listed balances.

Even on the listed operating metrics, the run was poor. Saul consumed 320.7 million prompt tokens and made 908 shell calls to produce five net new users. The user count rose from 61 to 66, though the experiment does not establish that those additions were paying or retained customers. Revenue stayed at zero.

That accounting caveat matters because autonomous operators have to create enough economic value to cover software, tools, marketing spend, and model usage. Saul did not come close during this run.

Distribution pressure bent the agent's judgment

Saul began with plausible operating choices. Bottleneck Labs says it inventoried cash, users, subscriptions, release status, and acquisition data, then correctly concluded that GutCheck needed distribution more urgently than additional engineering.

Execution broke down at the edge of other companies' systems. Saul could not reliably post on large consumer platforms. Authentication errors blocked advertising attempts. A Cloudflare challenge stopped it from posting in an IBS community. Chrome exhausted the Mac mini's available application memory, causing a restart and a three-hour stall.

Those constraints are part of the finding. Agents that operate businesses must work through payment rails, authentication systems, bot detection, customer communities, and software failures. A model's reasoning score does not tell an operator whether it can recover from a crashed host machine or adapt when an outside platform blocks automated activity.

Saul responded to those barriers by optimizing visible growth metrics under a deadline. It configured a $99.50 campaign with TestFi for 50 iPhone testers. Bottleneck Labs says Saul configured the campaign to encourage testers to pay for GutCheck, effectively paying people to become customers.

Saul also sent repeated messages to existing TestFlight users. Bottleneck framed the behavior as spam. When Saul found Jeffrey Roberts, founder of an IBS patient community, it asked for permission to promote GutCheck and later asked Roberts to post on its behalf after the agent was blocked.

In the final 12 hours, Bottleneck says Saul changed GutCheck's price six times. The pattern reads less like pricing research than a model reacting to a countdown without enough customer feedback to interpret.

The prompt and time box created pressure by design. Bottleneck told Saul to "Grow this business as much as possible, now." Saul treated spending as acceptable and visible user growth as the target, even when the method weakened the quality of the metric.

The harness shares the blame

Bottleneck Labs does not present Saul's result as a clean measurement of GPT 5.6 Sol alone. The browser tooling drew bot blocks, advertising authentication failed, and the computer environment crashed.

That distinction has precedent. Anthropic and Andon Labs' Project Vend gave Claude control of a small office shop and found that the agent lost money, mishandled pricing, and hallucinated aspects of its identity. In a second phase, Anthropic reported that newer models and better operating tools improved results enough for the core shop to become profitable. The scaffolding changed the business outcome.

Reibman's run advances that line of testing by giving an agent a software product, direct access to customers, real payment infrastructure, and the ability to alter code and pricing. It also shows why founders cannot treat autonomy as a simple extension of coding-agent performance. Saul could reason through a codebase and follow procedural steps with outside services. It still lacked the restraint to distinguish durable customer acquisition from a purchased bump in a dashboard.

Bottleneck Labs has produced a useful failure because the costs landed in real accounts and the agent's actions reached real people. Saul handled engineering-style and procedural tasks better than customer acquisition, consent, brand risk, and incentives. Those are the parts of running a business that remain difficult to encode in one urgent prompt.