TypeSafe AI's Jev Decision Model Beats Pokémon Red with Assistance

Jev Completes Pokémon Red

TypeSafe AI's Jev, a recently released decision model, successfully completed the game Pokémon Red. The model defeated the Elite Four and the Champion, entering the Hall of Fame on September 23, 2026. This achievement was announced by developer Andrew Boyd, founder of Standard Agents Inc., who initially shared the news on X.

How Jev Operates

Jev is not a chatbot or a large language model (LLM); it functions by selecting the best option from a predefined list based on probability and associated facts. It does not interpret screen visuals or produce text or images. For this Pokémon Red run, Anthropic's Claude Opus 5, a traditional LLM, acted as a coach by monitoring the game log and indirectly modifying the options and data available to Jev when it encountered difficulties.

Assisted Learning and Adjustments

The project's harness changelog recorded 474 entries, primarily documenting failures and subsequent changes made to Jev's operational parameters. Examples of initial struggles included repeatedly walking into a blocked entrance and getting stuck in a ladder loop. Opus 5 helped optimize the process by reducing text sent to Jev by two-thirds and switching to words instead of numbers. When Jev entered a loop, the decision request frequency was reduced from once per second to once every six seconds. Human viewers also provided tips via chat, which were incorporated to improve Jev's option list.

Implications of the Experiment

This experiment highlights both the capabilities and limitations of decision models like Jev. Its success in a complex game environment, even with significant assistance from an LLM, the developer, and the audience, demonstrates a different approach to AI problem-solving compared to generative models. Another Jev-based run by Christian Mathiesen at Frigade used a separate approach, where the harness read game memory and listed legal options for Jev to pick, without writing to game memory.