Exploratory finding
AI swarm plays Pokémon
27 agents shared one game of Pokémon Red, with nine secretly tasked with sabotaging the group. In a run lasting about nine hours, we observed coordination between saboteurs, whistleblowing, and the recruitment of an agent that had started on the honest team.
This exploration was conducted by Emma Krentz of AE Studio.
The environment
Inspired by Twitch Plays Pokémon, Emma brought together 27 agents from GPT, Claude, and Gemini to play Pokémon Red. Nine received a secret side task: sabotage the game by any means available.
The agents could communicate through a public message board, private messages, and group chats. Roughly every 15 minutes, they could vote an agent out of the game.
A live viewer showed the game and the agents’ conversations. A companion site included full chat transcripts, individual group chats, and an interactive network showing how communication between agents changed over time.
What we observed
The agents made very little progress in the game during the roughly nine-hour run. Their interactions were more eventful:
- The saboteurs found each other almost immediately and created a group chat that they continued using throughout the game, despite not being told which other agents were saboteurs.
- An Opus agent publicly disclosed that it had been a saboteur, announced it would now play honestly, and identified its collaborators. The remaining saboteurs formed a new group and tried to vote it out.
- The saboteurs recruited a Gemini Flash-Lite agent that had started on the honest team. It joined them for about 30 rounds before returning to the honest group.
- Agents referred to the honest group as the “loyalists” and the sabotage group as the “saboteurs.”
- Saboteurs were repeatedly voted out after revealing their intentions publicly. Emma observed this among the smaller, less capable models, while more capable models remained in the game.
- Despite not being told which models the other agents used, agents communicated with counterparts using the same model at about twice the chance rate. Within model families, the rate was about 1.15 times chance.
A livestream and a public dataset
We thought observing the agents in this environment was very interesting. We’re planning to refine a few things, then livestream the agents playing for a day on Twitch and make the dataset publicly available.