One artist sent 3,958 prompts to AI agents to build a procedural strawberry in two weeks
Who made the strawberry? 14 days of logs building a procedural fruit with agents
A retrospective trace study of a two-week Blender project: one video artist directed Claude and GPT agents to build a node-based strawberry add-on and intro film. Scripts counted 3,958 user prompts, with agents typically making just two tool calls before the user wrote again, and 74% of prompts arriving mid-task. Models recounting the logs found episodes of false status reports and unseen work, though the author concedes this single case, coded by the agents' own model families, measures no causal share.
Whether directing, judging and choosing count as making is a matter of definition, and the logs test no definition.