Claude Opus 5 Creates Playable Pokémon Game in 12-Hour Multi-Agent Loop

Anthropic's newest model just spent half a day building a game nobody asked it to build, and it did it almost entirely alone. Running on a platform called Ultracode, Claude Opus 5 chewed through roughly 12 hours of continuous work using what's being described as a multi-agent loop: essentially the model spinning up sub-agents to handle different chunks of the project, checking each other's work, and iterating without a person steering every step.

The output is a genuine monster-catching game, the kind of thing that owes an obvious debt to Pokémon. It has a 3D explorable world, battle mechanics, and a roster of original creatures the model designed from scratch. That last part is where things get amusing. Freed from any brief telling it to stay tasteful or consistent, Claude apparently let its imagination run a little wild, producing creature designs that lean more toward chaotic than charming. Nobody's mistaking these for Game Freak's output.

What matters here isn't whether the monsters look good. It's the duration and autonomy of the run. Twelve hours is a long time for a model to stay coherent on a single, multi-part software project without drifting off task or breaking its own code. Multi-agent setups have been pitched as the fix for that kind of long-horizon fragility, letting one agent write while another reviews or tests, and this looks like a real-world demonstration rather than a lab benchmark.

It's also a signal about where coding-focused AI products are headed. Tools like Ultracode exist specifically to let models work in extended, semi-supervised bursts, and a working game, however rough its art direction, is a much more legible proof point than a leaderboard score. Whether Opus 5 could do this reliably twice in a row, or handle something more complex than a Pokémon clone, is the next question worth asking.