Give a coding agent a design team
Developer Chukwuduzie Blaise built Skulptor, an open-source skill that separates frontend work among a designer, builder, and critic. Kimi K3 turns a request and reference images into a prototype, DeepSeek implements it, and Kimi compares screenshots with the prototype.
In Blaise’s bakery-page experiment, that handoff produced an almost pixel-for-pixel implementation, but the design still felt safe and left too much empty desktop space. The lesson is useful beyond these models: separating roles can improve execution without supplying taste.
Let an agent explore before assigning the task
Researchers placed copies of Claude Opus 4.7 on a simulated island for 30 hours and let them explore without a specific task. The agents created drawings and games, including versions of basketball and the floor is lava.
Agents with that experience later performed better than fresh copies on assigned island tasks, while deleting lines from their memory files reduced performance. The paper has not been peer-reviewed, and outside researchers said the behavior alone does not establish that the agents were genuinely playing.
Also
- Prepare cleaner reference images for Firefly. Alice Severi Gonçalves reports that removing backgrounds and standardizing her dog photos to square canvases stopped surrounding objects from leaking into generated illustrations.
- Why mixed-language questions can trip up models. Adobe researchers found that switching languages often reduced question-answering performance and propose an inference-time intervention to recover accuracy.
- Character.AI follow-up: Disney offers jobs. After the bankruptcy covered October 13, a legal filing says Disney agreed to offer jobs to at least 60 employees, and most recipients accepted.
Pass it on