AI agents built the olddogs.ai website — the full process was shown

olddogs.ai author Ben Kamens explained how the company's website and the article about its creation were written by Claude Code and Codex agents, and the entire process — from 34 generated 'visual worlds' to a renderer rewritten in a GPU shader (14 → 117 fps) — was published in an open archive along with 2,460 intermediate images.

Compiled by Sergey KostenchukPublished 2026-10-02Updated 2026-10-02

2026-10-02 Coding Anthropic

🛠 AI agents built the olddogs.ai website — the full process was shown

Ben Kamens explained how the website and article were written by Claude Code and Codex agents, and all 2,460 intermediate images were uploaded to an archive. Design was found through generation, not code: 34 'visual worlds' before layout, then the variant was checked with screenshots against a mockup. The Hero Lab panel exported JSON back to the agent, and the request 'make it much faster' rewrote the renderer in a GPU shader: 14 → 117 fps, plus 77 KB.

🌍 Quality is determined by the process, not by luck: sketches before code, checking screenshots against mockups, settings return to the agent in a single JSON block. Essentially, this is a human-in-the-loop eval — at each step, a human selects.

👤 Working with Claude Code or Codex — the techniques are repeatable now: the article has ready-made prompts, JSON Hero Lab, and a mosaic of all 2,460 images. One successful shader is still n=1.

Source 1: https://olddogs.ai/blog/how-we-made-olddogs-ai-with-ai

Sources