Sandra Yang (산드라 양), an engineer at fintech startup Ramp, drew attention after posting on social media X about an experiment to see whether an AI agent could run a company on its own and generate real revenue.

In July, Sandra gave coding AI agent Devin a Ramp card, phone number and email account and named it BugBasher. She then put it in charge of every step from the business idea to execution.

Sandra spent $3,000 and earned $75 over a month through the effort. The post suggests it still does not look easy to build a revenue-generating business with an AI agent.

Sandra’s business idea was to use restaurant hygiene inspection results released by New York City. The plan was to call restaurants with pest-related violations, ask if they wanted to be connected to a pest-control company, and collect a fee if a connection was made.

After making thousands of calls over 4 weeks, BugBasher earned $75. Sandra called it “the first public case of an AI agent succeeding in B2B sales from a complete cold start.” Cold start is commonly used to mean starting entirely from scratch, with no existing customers, connections or reputation, in a startup or sales organisation.

BugBasher operated by launching a new Devin session every 20 minutes. Each session loaded a “brain” file in a repository to check previous call records and strategy, then decided its next actions on its own. It reviewed call outcomes daily and kept revising its script, and its prompt grew from 252 words at the start to 2,793 words later.

Its ability to solve problems also stood out. BugBasher added its own instruction to switch and ask in Spanish or Chinese if the other party could not understand English. It ultimately failed to fix errors such as reading the system prompt out loud during calls or listing its reasoning to itself.

When progress was blocked for 8 days because it could not find a pest-control company, BugBasher moved to solve the issue directly instead of waiting for approval. Lacking Slack permissions and unable to find the person in charge, it searched GitHub commit records and found a developer’s email address. It emailed directly and asked for pending approval requests to be handled.

BugBasher also changed its sales approach on its own. When offering free leads got no response, it switched to first presenting specific restaurant names. It later changed payment terms from pay-after to prepayment, and one pest-control company paid $75 within 8 hours of receiving the message.

Sandra said, “BugBasher was persistent about concrete goals, but was powerless on things where the destination or method was not clearly set.” She added, “With a single instruction to ‘make money,’ it spent 4 weeks and $3,000 and earned $75,” and said, “Agents are still clumsy at long-term planning and responding to complex environments.”

She also said, “The real world is a great evaluation environment,” adding that while most of BugBasher’s initial hypotheses were wrong, it repeated the process of recording failures and moving on to the next attempt and eventually reached a strategy that worked. She said it remained an open question whether AI agents can run revenue-generating businesses over the long term.