The sub-agent mode generated a museum heist with raccoon stacking, though it missed a floating eyeball bug.
Simon Willison reports that Codex Desktop running GPT-5.6 Sol Ultra produced a much better Raccoon Heist game than Claude Fable 5 did from the same one-shot prompt. The new game, Moonlight & Mayhem, places the raccoons in a museum heist to rescue crewmates and stack on top of each other.
Willison previously had Claude Fable 5 build a simpler game with a single raccoon in a backyard. The original premise came from GPT-3 and DALL-E four years ago. Codex spent 52 minutes on the project and generated textures and prompts using gpt-image-2.
The test shows that aggressive sub-agent use can generate richer game concepts and code, but the process still misses obvious visual flaws. Codex reviewed screenshots during development and failed to spot an eyeball enlarged to a giant sphere over each raccoon. Willison fixed the bug by prompting the model to explain the spheres and then to fix them.
Willison shared the full Codex transcript in the repository and noted he wishes Claude Code had the same copy-as-Markdown feature. He also included an AgentsView cost estimate for the session at full API prices. The repository contains the textures and prompts that Codex generated.
What matters
- Codex with GPT-5.6 Sol Ultra built the Raccoon Heist game in 52 minutes from a one-shot prompt.
- Aggressive sub-agent use improved code generation but missed a prominent visual bug in review.
- Simon fixed the eyeball bug with direct prompts and shared the full Codex transcript on GitHub.
Why it matters
Simon fixed the eyeball bug with direct prompts and shared the full Codex transcript on GitHub.
This GenAI News article was prepared in original wording using reporting and materials published by Simon Willison’s Weblog. Source reference: https://simonwillison.net/2026/Aug/7/moonlight-mayhem/#atom-everything.
Drafted by the GenAI News review pipeline.