A source-code breakdown of Astra's computer use loop
The author read the source code to analyze the agent loop behind Astra's ability to operate Blender and generate 3D worlds.
What this article says
文章从 GPT-6 Astra 用 Blender 建模、生成 3D 世界这一现象切入,指出这些能力其实是它新增的 computer use 能力的副产品。作者表示自己直接阅读了相关源码,重点拆解让 Astra 表现突出的那个执行循环:模型如何观察屏幕、决定下一步操作、调用工具并在多轮之间维持状态。结论是 Astra 的强项不在于单次生成质量,而在于这个循环把视觉输入、动作执行和反馈校正串成了可稳定重复的流程,因此才能在 Blender 这类复杂 GUI 软件里连续完成多步任务。文章面向想理解 agent 与 computer use 实现细节的读者。
GPT-6 Astra broke the internet with its ability to use blender and generate 3D worlds, all a byproduct of it's new computer use capabilities.
I dove into the source code and wrote about the loop that makes Astra so good: https://x.com/i/article/2097058174125944832