Saying that "AI is good at coding" is no longer something to brag about. It's just a fact. The more interesting question is something else. "Can an agent beat Pokémon?"
This question actually gets to the heart of LLM agents. Writing a few lines of code is easy. But looking at the screen, remembering its current location, not losing sight of its goal, and, when it runs into a wall, saying, "Oh, that's not it," and finding the way again. Current AIs are surprisingly bad at this absurdly simple loop.
The point this event wants to show is not "how smart AI is." Give Pokémon to pretty much any model, and it will loop endlessly in front of the same wall. It can't reliably remember either its current location or its next goal.
So here's the challenge: "Is there a device that can get a failed AI back on target?"
Observe, try again, check the goal. The harness is what ties this loop together. It's the work of combining items like save state, memory, retry, and goal check to force a lost agent back to life. Instead of grandly talking about an "LLM agent benchmark," it's simply an event where everyone plays Pokémon together and watches the AI lose its mind.
The goal is to reach the Viridian City section in Pokémon Red. Tear apart, fix, and bring a Pokémon game agent harness back to life. Whether you make it play Pokémon honestly or create a crazy pathfinding monster, the method is up to you. All that matters is that the agent reaches Viridian City autonomously.
Participants enter like Pokémon trainers. They observe the AI wandering around (FIGHT), record failure patterns like entries in a Pokédex (DEX), take out and assemble harness items (BAG), and try again with the modified harness (RUN). At the end, everyone shares their progress logs and funniest moments.
What do you get by participating? To be honest: agent-debugging PTSD, building a hellish state-management system, the realization that humans are more robust than you thought, and, if you're lucky, an AI that actually beats Pokémon. Everything except the last item is guaranteed.
ULWmon : Harness Hackathon will take place on Saturday, June 6, from 11 a.m. to 3 p.m. at VYV House in Cheongdam-dong. Registration is confirmed after host approval. Backed by Hashed.
PokemonDev : AI가 포켓몬을 깰 수 있을까?포켓몬 레드에서 초록마을까지. 에이전트 하네스를 깎고 AI가 정신 나가는 걸 같이 봅니다.pokemondev.comOpen
