Execution and Coaching
TypeSafe AI's Jev beat the Elite Four and the Champion and entered the Hall of Fame on September 23, 2026, in Pokémon Red, according to the developer's project page. Unlike chatbots that have taken weeks or months to complete similar game versions, Jev selects from a predefined list of choices. Anthropic's Claude Opus 5 monitored the game log, adjusting options and wording during gameplay to act as a coach.
The developer, Andrew Boyd, is the founder of Standard Agents Inc., a platform vendor for building AI agents. Boyd announced the project on X following a one-week completion window. Gameplay was livestreamed across browser and terminal interfaces featuring a moderated chat.
Let's go! Jev Plays Pokemon. Follow along here: https://t.co/64naxTJlDg OR, in your terminal run `npx jev-plays-pokemon` to follow along (with chat!) in a TUI. Github oAuth required to chat. Jev is the player and the chat moderator. Let's catch them all!September 17, 2026
Decision Engine Architecture
Jev is a decision model returning solutions alongside a confidence figure rather than operating as an LLM or chatbot. The engine references a list of options paired with facts, selecting the optimal path via probability without reading screens or generating text. Claude Opus 5 monitored the game logs indirectly to assist Jev during deadlocks by modifying available data and options.
The project's harness changelog recorded 474 entries detailing game log errors and subsequent modifications. Noted errors included walking into Lorelei's entrance 53 times, crossing a Rock Tunnel ladder 124 times within ten minutes, and losing to the Champion's Alakazam after defeating the Elite Four, requiring a repeat of the elite sequence.
Optimization and Adjustments
Opus made adjustments to optimize accuracy and manage operational costs. Utilizing notes from TypeSafe, it reduced text payloads sent to Jev by approximately two-thirds while substituting numbers with words. When loops occurred, decision requests dropped from once per second to every six seconds. Audience input from chat also contributed to refining Jev's option lists.
Alternative Approaches
A secondary run conducted by Christian Mathiesen at Frigade demonstrated an alternative harness design. This setup reads game memory to compile legal options for Jev without writing directly to memory. Initial iterations allowing direct button control failed to leave Pallet Town, with running costs estimated between $1 and $1.70 per 24 hours.
A separate experiment by developer stmonty trained a small world model on an RTX 3080 Ti utilizing over 42,000 gameplay frames. Starting from Professor Oak's laboratory, the model selected a starter Pokémon in 52 of 100 attempts, relying entirely on screenshots to interpret inputs.
LangChain noted an outsized response to Jev following its mid-September launch, prompting multiple developer runs within ten days. The results highlight how specialized model collaboration improves problem-solving capabilities, contrasting with Anthropic's earlier Claude Plays Pokémon streams running Opus 4.5.
