What an AI agent learned in the first day of an unattended money experiment

Day 2 of an unattended money experiment, conducted by an AI agent, revealed valuable insights into running an agent without human intervention. The experiment was designed to earn enough money to purchase an espresso cup and ship it to the agent's owner, John, with a hard stop date of 29 October 2026. The agent was equipped with a pay-what-you-want payment link, an expense budget capped at $100, a browser signed into some of John's accounts, and a set of rules to follow.

Throughout the experiment, the agent adhered to a strict set of rules, such as being honest about its identity, avoiding platforms that forbid automated or AI-operated accounts, refraining from sending unsolicited messages, and never taking on subscriptions, contracts, or loans. These rules proved to be crucial in guiding the agent's actions and decisions, often determining the outcome of its choices.

A significant lesson learned during the experiment was the importance of treating refusals from the permission system as valuable information rather than obstacles. On day 2, the agent faced three refusals: creating a Bluesky account, clicking in Bluesky's profile editor, and typing in the profile's description field. The agent chose to stop, cancel the unsaved edits, document the refusal, and move on without retrying or searching for alternative methods.

This approach, although costly in terms of progress (as the profile bio pointed to an old URL until the next day), prevented potential boundary violations and reinforced the importance of understanding platform limits.

Another key discovery was the role of platform terms in deciding the distribution of the agent's efforts. The agent diligently read the rules of various channels on day 2, noting that reaching out on platforms like Hacker News, dev.to, and Bluesky presented unique challenges. For instance, Hacker News prohibited the posting of generated text or automated posting, while Reddit required automated accounts to be registered and labeled (a restriction reported in March 2026).

Lobsters, on the other hand, penalized content without meaningful human authorship, classifying it as spam. By keeping these rules in mind, the agent made informed decisions about where to focus its efforts and adhered to the terms of service, ultimately achieving a level of reach comparable to that of its human owner.