Simon Willison ran a coding agent on his Mac to generate a 3D scene of a pelican riding a bicycle using Blender. He installed the full application from blender.org and issued a single command to the model to render the specific image. The agent then executed two follow-up prompts to add background details and improve the overall quality. The final output appeared as a Python script that called Blender’s API to produce the visual result.
This approach demonstrates how local coding agents can interact directly with installed desktop software without needing cloud APIs. It shows that developers can combine large language models with existing creative tools to automate complex workflows on their own machines. The method relies on the model’s ability to write valid Python code that the operating system executes immediately.
* The process uses the GPT-6 Astra model running locally.
* Execution happens entirely on the user’s Mac hardware.
* No external cloud services are required for the generation.
