Joel Z

ALT Flowchart of the Gemini Plays Pokémon "Vision First" harness. Top row: the node-mgba emulator, shown as a pixel-art game screen, sends the frame and player (x, y) to the harness and gets button inputs back. The harness, pictured with a clean screenshot, turn filmstrips and a coordinate grid, sends screen and context to Gemini, which returns tool calls. Middle: the agent turn lifecycle in six steps: capture, observe, save assets, build context, reason, execute, then loop to the next turn. Memory upkeep: reflect every 20 turns, critique every 30, summarize about every 125. Bottom: four tool groups. Gameplay: press_buttons, speak, crop, update_objectives. Notepads: edit, read, load, unload, search, delete. Python sandbox: run_code plus file read, write, edit, delete. Sub-agents: invoke, define, delete, new_context, report_outcome. A note says the overlay minimap and party HUD are for viewers only and never reach Gemini.