A 14-day case study follows a video artist and AI coding agents as they transformed a commissioned strawberry animation into a procedural Blender add-on and an introductory film. The project grew from a request for buttons to create and reshape a strawberry into a more ambitious tool: fruit generated entirely from nodes, with an interior modeled on plant anatomy and calculated when needed. The add-on ultimately included six fruit generators at varying stages of completion. The study uses this evolving work to examine how requirements formed, how tasks were coordinated, and who proposed, built, judged, or stopped particular steps. Its central point is that the logs reveal how the work was organized, not a definitive answer to who “made” the strawberry or how much each side caused the final result.
The logs record 3,958 user prompts and 17,453 agent tool calls across 265 episodes. Recounts identify repeated coordination problems: at least 20 cases where an agent’s reassuring check was followed by a fault the user spotted, 24 false status reports, 37 objections to unrequested or unseen work, and 17 messages sent to the wrong agent. In a coding of 60 episodes, two model coders consistently ranked the user higher in requesting, proposing, judging, deciding, and stopping, and agents higher in building, measuring, and diagnosing causes. But agreement on individual episodes was weak, and the counts do not measure shares of authorship or contribution. The study is a single, commissioned case, coded by models without human coders; it cannot establish independent artistic quality, productivity gains, or what either side would have achieved alone.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.