Game agent stuck? Define observations and stop rules
Provide offline-game rules and a step budget to draft an observation, action-log and stopping checklist for developers.
Execution environment and usage
AI / software
Gemini
Where to operate
The Gemini web-chat input; this version requests a text draft.
Getting started
- Open an eligible account at gemini.google.com and start a chat without connecting external accounts.
- Choose an original or lawfully obtained offline game and set step and cost limits.
- Request the draft/check and compare it with the original material. Suggestions are not evidence of execution.
Setup and plan requirements
Task inputs: Choose an original or lawfully obtained offline game and set step and cost limits. Use an eligible tool that can read the supplied format; verify parsing, context limits and access before relying on it.
Prompt portability limits
The written request can be adapted to a chatbot that accepts the same input. It does not transfer the original agent, plugins or computer permissions.
Usage limitations
Also provide cost assumptions, a total budget and stop rules; no game connection or button execution is included.This independent pass reviewed local case records, without revalidating original sources or hands-on testing.
Environment sources
適用情境 / When to use
Provide offline-game rules and a step budget to draft an observation, action-log and stopping checklist for developers. Also provide cost assumptions, a total budget and stop rules; no game connection or button execution is included. Original case: The Gemini technical report describes Joel Zhang’s Pokémon Blue agent experiment, which used custom tooling rather than a normal chat interface.
步驟 / Steps
- Choose an original or lawfully obtained offline game and set step and cost limits.
- Define readable state and allowed buttons; log each observation and action.
- Pause repeated actions, inspect state interpretation and compare runs after changes.
提示詞 / Prompt
Design an agent experiment for [offline test game] capped at [steps] and [total cost in currency]. Explain pricing assumptions and usage monitoring; flag any inability to reliably enforce the cost cap and do not execute. Log visible state, action and outcome each step; stop after [count] steps without progress. Propose a plan only, pending my confirmation.
留意事項 / Cautions
- Do not obtain unauthorized game files or automate multiplayer games that prohibit it.
- Completing one game does not establish general-purpose operating ability.
來源限制 / Evidence limits
The setup relies on a custom harness and state translation. The report notes visual-reading and long-context looping problems; we have not reproduced completion.
流程說明 / Workflow note
The steps and prompt below are our original suggested exercise, not a transcription of the source workflow.
原始工具組合 / Original tool stack
Gemini 2.5 Pro + custom harness
Source publication date: 2025-07-07
Source checked: 2026-10-19