Moonlight & Mayhem (Raccoon Heist by Codex + GPT-5.6 Sol Ultra)
Simon Willison’s Weblog Simon Willison ● Covered by 2 sources
Simon Willison had two AI coding agents build the same raccoon heist game from an old prompt. Codex's version was way better — except every raccoon grew a giant floating eyeball.
Based on reporting by Simon Willison’s Weblog, Simon Willison — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Simon Willison likes running the same prompt through different AI tools just to watch them diverge. Earlier in the week he had Claude Fable 5 turn a four-year-old GPT-3-and-DALL-E game pitch into a working browser game, a cute little thing where a lone raccoon wanders a backyard scooping up coins and fish. This time he handed the identical premise to Codex Desktop running GPT-5.6 Sol Ultra, the configuration where the model leans hard on sub-agents to split up the work, and the result was a different beast entirely.
The original pitch described a crew of thieving raccoons pulling off bank jobs and art heists, and Sol actually built that. The finished game, called Moonlight & Mayhem, drops you into a museum where you have to free two raccoon crewmates and physically stack them on top of each other to reach a golden sardine locked in a display case. That's a real heist mechanic, not a scavenger hunt, and it's a much closer match to what the four-year-old prompt actually asked for.
It wasn't flawless. The first one-shot build shipped with a glitch where each raccoon sprouted an eyeball the size of a bowling ball, floating over its head like some kind of cursed halo. Codex had reviewed its own screenshots during development and somehow missed it entirely. Willison fixed it the low-effort way, just asking the agent why the raccoons had huge black spheres attached to them, then telling it to fix it — and that one-line nudge did the job.
The build itself took 52 minutes, sub-agents and all, and Willison published the full transcript in the GitHub repo along with the textures and prompts Sol generated using gpt-image-2. He noted the transcript exports cleanly as Markdown, something he wishes Claude Code supported. He also ran the numbers on what that session would have cost at standard API rates rather than through his flat-rate Codex subscription, and it's a useful reminder that these agentic coding sessions, sub-agents included, are not free even when the subscription makes them feel that way.
My take — AI-written commentary, not fact-checked reporting
The gap between these two builds says more about agent architecture than raw model smarts — sub-agents apparently buy you a game that actually matches the brief instead of a watered-down version of it. But the eyeball bug is the real story: an agent that can spin up a museum heist with physics-based stacking still can't glance at its own screenshot and notice a giant black sphere hovering over a raccoon's head. That mismatch between capability and basic visual sanity-checking is exactly the kind of thing that should worry anyone handing these tools more autonomy before they can reliably grade their own homework.
Read more about this at: Simon Willison’s Weblog
Related stories
Speed Racer Browser Game Achieves 18+ Hours of Opus 5 Iteration
speed-racer-ten.vercel.app · 5 days ago ·
46
Claude Opus 5 Creates Playable Pokémon Game in 12-Hour Multi-Agent Loop
Reddit · 5 days ago ·
28