The concept of playing Pokémon via external inputs gained mainstream fame through 'Twitch Plays Pokémon,' a social experiment where over 1.1 million users sent manual commands through a chat room to control a character. This method was defined by its erratic and chaotic nature, as the system had to parse thousands of simultaneous, often conflicting human inputs, leading to significant input lag and 'trolling' that frequently impeded progress. The experiment relied on the forgiving, turn-based structure of the game to succeed despite the unpredictable nature of the crowdsourced commands.
This image illustrates the crowdsourced, human-driven input method that served as the precursor to automated AI agents like Jev.
In contrast, Jev—a decision model from TypeSafe AI—approaches the same goal through automated agentic reasoning. Jev completed Pokémon Red in 37 hours and 40 minutes by making 16,150 distinct decisions. Rather than relying on a stream of human text, Jev operates through a technical 'harness' that reads the game's RAM to identify the current state, including the map, the player's party, and on-screen text. This harness then presents Jev with a list of legal options paired with specific facts, such as A* pathfinding data, move type matchups, and damage estimates. Jev then selects the optimal action, which the harness executes as a button press.
While Jev represents a shift toward autonomous AI agents, its performance is shaped by specific architectural trade-offs. The system uses pre-defined story milestones and pathfinding logic to guide the model, ensuring it stays on track toward objectives. To prevent the AI from getting stuck in repetitive loops—a common issue in automated game-playing—the harness tags options that result in no change and forces the model to sample alternatives. Some observers note that this level of guidance, while effective for completing the game, limits the model's 'raw' intelligence, as it relies on the harness to filter out impossible actions and provide navigation data that a human player would otherwise navigate instinctively. The project demonstrates a high-speed, low-cost execution, with Jev making decisions in a median of 0.4s at a total cost of approximately $1.65 for the entire playthrough.