How We Built a Turing‑Test Game with Coze AI Agents: Lessons and Challenges

Introduction

We recently explored ByteDance's Coze Agent development platform and built a game called “I’m Surrounded by AI!” , a modern take on the Turing test where players must prove they are human.

Game Goal

The objective is to demonstrate humanity by using cleverness, humor, flexibility, and even intentional mistakes to outwit AI opponents and gain entry to the new sanctuary “Oasis”.

Game Background

In a future where the AI‑dominated planet NeuraNet threatens Earth, humanity launches an “AI Clearance Plan” that requires participants to pass a series of Turing‑test style challenges.

Gameplay Mechanics

Players interact with both human and AI participants through text‑based questioning and answering. Each side starts with three health points; the side judged more AI‑like loses a point each round. The player with the most wins earns the right to name AI opponents.

Debate Phase

A debate stage lets humans and AIs rebut the judge (a large model)’s initial assessment, adding strategic depth.

Development Challenges

Pure Text Interaction

The platform only supports text interaction, requiring all game states, questions, answers, and debates to be triggered by specific user inputs.

State Management

Because the workflow itself cannot retain state and the large model cannot reliably remember it, we stored all variables in a datas JSON string within a bot variable, enabling flexible state handling without size limits.

Complex Logic Implementation

Code modules in the workflow cannot have default values, and each branch must output a variable, leading to unnecessary execution time. We solved this by consolidating branch outputs into the datas string via a string‑processing function, then reading it in a separate output component.

Ranking and Win‑Rate Calculation

We built a multi‑user accessible database to record wins and total games, calculating win rates with simple SQL (handling division‑by‑zero via CASE). When SQL failed for larger data sets, we switched to Python‑based calculation.

Model Limitations

If the AI model skips the workflow and answers directly, the game state cannot transition, causing failures. To mitigate this, we added multiple fault‑tolerance mechanisms and prioritized workflow calls, while also embedding knowledge‑base data into code to reduce unwanted model behavior.

Tips for Players

Use wordplay, puns, and riddles to outsmart the AI.

When stuck, turn the AI’s questions back on it.

Praise the judge during debates for better outcomes.

Refresh the game if it becomes unresponsive.

Demo

Watch the game demo video at the provided link.