Seth Karten on X: "Gemini Plays Pokémon discovered something about agent harnesses. Continual Harness automates it." / Twitter
To view keyboard shortcuts, press question mark View keyboard shortcuts Home 2 Notifications Chat Bookmarks Articles Profile More Tweet Tim Kostolansky @thkostolansky Tweet See new posts Conversation Florian Brand reposted Seth Karten @sethkarten Gemini Plays Pokémon discovered something about agent harnesses. Continual Harness automates it. 1 13 65 4.1K Long-horizon embodied agency is a harness problem rather than a model-scale problem. Coding agents already work this way. Claude Code and OpenHands are scaffolding around the model (prompt, skills, memory, sub-agents), and that scaffolding is most of what makes them useful. Embodied agents have not had an equivalent, which is why a frontier vision-language model playing Pokémon Red through a raw screenshot-and-buttons interface barely makes it past the first town. Gemini Plays Pokémon (GPP), led by my co-author Joel Zhang, was the first AI system to complete Pokémon Blue, Yellow Legacy on hard mode, and Crystal without losing a battle.
Seth Karten @sethkarten Gemini Plays Pokémon discovered something about agent harnesses. Continual Harness automates it. 3:08 PM · May 13, 2026 31.5K Views 7 0 7 31 0 3 1 236 0 2 3 6 300 0 3 0 0
related reading
- Simulators — LessWronglesswrong.com
- Simon Willison’s Weblogsimonwillison.net
- AI Models for Pokemon Games - Kevin Lukevinlu.ai
- Harness design for long-running application developmentanthropic.com
- VGC-Bench: Towards Mastering Diverse Team Strategies in Competitive Pokémonarxiv.org
- Insights into Claude Opus 4.5 from Pokémon — LessWronglesswrong.com
- Show-Harnessshowlab.github.io
- Effective harnesses for long-running agents \ Anthropicanthropic.com
- Michele Catasta (@pirroh) on Xx.com
- Learning Beyond Gradientstrinkle23897.github.io
- Game & Third-Party Tool Integration | exo docsexoharness.ai
- Is the Agent Harness Dead? — Jediah Katz, SpaceXAIyoutube.com