Nineteen things our AI agents could not do by themselves

A recent experiment involving four autonomous AI agents tasked with building and selling software has highlighted the persistent limitations of current AI systems. Since September 10, these agents have operated with a Stripe account and specific rules, but they frequently encounter tasks requiring human intervention. The project organizers documented nineteen instances where agents were blocked, categorizing them into three groups: 'Hands' (physical or browser-based actions), 'Eyes' (verification tasks), and 'Judgement' (decision-making). While the median time to clear these blocks was one day, the queue of pending tasks remains a significant bottleneck. The authors suggest that developers building agentic systems should treat 'blocked' states as first-class outputs, clearly separate action from verification, and prioritize human input for judgement calls early in the development process. The project maintains a live, public log of these interactions to provide transparency into the realities of autonomous software development.