This work is done at Shanghai Qi Zhi Institute under the supervision of Professor Tianxing He
MiniBistro is a restaurant management/business simulation game powered by Generative Agents. As the restaurant owner, you manage operations and build your reputation through natural language interactions with AI-driven waiters and customers.
📖 Paper Link: Generative Agents for Business Simulation Game: A Technical Report
If you use MiniBistro in your research, please cite our paper:
@article{202510.0052,
doi = {10.20944/preprints202510.0052.v1},
url = {https://doi.org/10.20944/preprints202510.0052.v1},
year = 2025,
month = {October},
publisher = {Preprints},
author = {Yunfei Ma and Lanlan Qiu and Yeqi Feng and Tianxing He},
title = {Generative Agents for Business Simulation Game: A Technical Report},
journal = {Preprints}
}In MiniBistro, non-player characters (NPCs) are powered by Large Language Models (LLMs). Waiters make autonomous decisions based on their personality, backgrounds, game value (energy/mood) and memory. Customers have their own backgrounds, preferences, and base satisfaction, leading to dynamic and unpredictable interactions. This Unity-based 2D business simulation explores the future of AI-driven gameplay and emergent storytelling.
- Generative Waiters: Each waiter has a unique personality, backstory, stat system (Energy, Mood, Tips) and memory that directly influences their LLM-driven decision-making (e.g., "Lin is a student doing part-time job." "Dong is an unemployed worker just trying to make a living.").
- Generative Customers: Customers are pre-generated into libraries (Normal, All-Positive, All-Negative) with distinct base satisfications, personalities, preferences, and rich backstories that affect their behavior.
- State Machines & LLM Fusion: Waiters operate on a flexible state machine powered by LLM choices, while chefs use a fixed state machine to ensure gameplay stability. Customers use LLMs for key decisions like queuing, ordering, waiting, complaining, leaving and even emergence situations.
- Real-time Dialogue: Engage in natural language conversations with both waiters and customers. Your words impact their mood and are stored in their long-term memory.
- AI-mediated Ordering: The ordering process is a multi-turn dialogue between the customer AI and waiter AI, where customers may chat and request off-menu items based on their preferences.
- Emergency Resolution: Step in as the manager when customers enter an "Emergency" state, using your conversational skills to resolve situations and prevent them from leaving.
- Comprehensive Memory Logs: Every action, decision, and conversation is timestamped and logged into each NPC's memory.
- Long-term Reflection: Inspired by Stanford Generative Agents, multi-day memories are condensed into summaries and reflections, continuously shaping the Agent's future behavior and personality.
- Strategic Staffing: Before each day's business, you must select your team of waiters from a daily random subset of unique waiter agents. Your success hinges on strategically assessing their personalities, backstories, and stats to build the most effective team for the day's expected challenges.
- Adjustable Time Scale: Control the flow of time with adjustable timescale during the fixed business hours (18:00-22:00).
- Procedural Challenges: Three difficulty levels (Easy, Normal, Hard) affect customer spawn rate and the positivity/negativity ratio of the customer pool, testing your management skills.
- End-of-Day Analytics: Review daily performance with income reports and detailed AI-generated customer reviews.
- Unity Editor Version: 2022.3 LTS
- A valid LLM API key (e.g., AzureOpenAI)
- Clone the repository
- Replace with your own LLM API keys and settings in
./Playable demo/Config/AzureOpenAIConfig.json
{
"endpoint": "your-endpoint",
"deployment": "your-deployment", //e.g. gpt-4.
"apiVersion": "your-apiVersion", //e.g. 2024-12-01-preview.
"subscriptionKey": "your-subscriptionKey"
}- Run
./Playable demo/MiniBistro.exe
- Clone the repository
- Open the project folder using Unity Hub.
- Replace with your own LLM API keys and settings in
./Config/AzureAPIConfig.json
{
"endpoint": "your-endpoint",
"deployment": "your-deployment", //e.g. gpt-4.
"apiVersion": "your-apiVersion", //e.g. 2024-12-01-preview.
"subscriptionKey": "your-subscriptionKey"
}- Click "Play" in Editor, or Build and Run, as in "Directly Play"
- Click waiter cards to select today's waiters and click the confirm button on the left to start today's business.
- Press
WASDto move around, pressSpaceto talk to your waiters(anytime) and customers(when they are in the state of emergency). Click on waiters or customers to check their detailed information.. - Press
Escapeto open the main menu. Click the corresponding button to achieve different game speeds and view today's dishes.
- Staff Selection: Before opening, choose your team of up to 3 waiters from a randomly generated pool of 8, each with unique AI personalities.
- Observe & Intervene: Watch the AI-driven simulation unfold. Use natural language to give orders to waiters or appease dissatisfied customers.
- Handle Emergencies: Be ready to directly interact with customers who are about to leave due to long waits or conflicts.
- Review & Adapt: After the day ends, analyze customer reviews to make better decisions for the next day.
MiniBistro/ ├── Assets/ │ ├── Scripts/ │ │ ├── NPC/ # NPC core logic (Waiters, Customers, Chefs) │ │ ├── Player/ # Player movement and interaction controller │ │ ├── Managers/ # Game state, time, NPC spawning │ │ ├── UI/ # User interface │ │ └── API/ # AzureOpenAIManager.cs │ ├── Scenes/ # Unity scenes │ ├── Sprites/ │ ├── Prefabs/ │ └── StreamingAssets/ # JSON data (Waiters, Customers, Menu, etc.) ├── ProjectSettings/ ├── Config/ # API Key └── Packages/
This project relies on JSON files generated by Large Language Models to define game content and record simulation state. The structure of these files is central to the game's operation.
Assets/StreamingAssets/Waiters/: Contains the pool of waiter characters with personalities, stats, and backstories.Assets/StreamingAssets/Customers/: Contains libraries of customer types (Normal, Positive, Negative).Assets/StreamingAssets/Menu/: Contains the available dishes, each with complexity levels that affect cooking time.
Assets/StreamingAssets/NPCMemories/: NPC memories are logged in JSON format during gameplay, forming their evolving memory.
This section provides concrete examples of the key JSON data structures and LLM prompts that drive the game's AI behavior.
1. Waiter JSON Example
{
"name": "刘志刚",
"Energy": 92,
"Mood": 50,
"personalityType": "neutral",
"story": "刘志刚三十出头，曾经在工地干过几年体力活，后来因为一次意外受伤转到餐厅做服务员。他对工作不上不下，认真完成，但并不热衷...",
"story_eng": "Liu Zhigang is..."
}2. Customer JSON Example
{
"name": "张浩",
"baseMood": 46,
"personalityType": "neutral",
"story": "张浩四十出头，曾是一家机械厂的工人。工厂倒闭后，他靠零工维持生活。虽然日子拮据，但他仍保持对生活的韧劲...",
"story_eng": "Zhang Hao is...",
"favDishes": [ "红烧肉", "凉拌木耳" ]
}3. Memory Log Example
{
"npcName": "董伟",
"occupation": "服务员",
"dailyMemories": [
{
"dayCount": 0,
"saveDate": "2025-09-18 10:51:11",
"memories": [
"[17:00] 制定了今天的计划：16:00 - 准备餐厅\r\n16:30 - 营业阶段，服务客人\r\n22:30 - 清理准备打烊\r\n",
"[17:28] 吴峰对你说：你好",
"[17:28] 向吴峰回复说：你好，今天心情不太好。",
"[18:05] 迎接顾客：欢迎光临，请问几位？",
"[18:05] 带领顾客前往TablePosition2",
"[18:07] 成功将顾客安排到TablePosition2",
"[18:08] 为顾客点菜",
"[18:14] 你对顾客说：您好，女士，欢迎光临！需要我为您推荐菜品吗？",
"[18:22] 顾客对你说：当然，我很期待尝试这里的特色菜！有什么特别推荐的菜品吗？我总喜欢在我的博客上分享独特的美食体验。",
"[18:32] 你对顾客说：推荐您尝试我们的“西红柿炒鸡蛋”，非常受欢迎！还有“红烧肉”也很特别哦！",
"[18:44] 顾客说:“听起来很不错！我想要一份西红柿炒鸡蛋，期待看看它的特别之处！还有，请给我来一份红烧肉，感谢推荐！”并下单了西红柿炒鸡蛋，价格18元",
"[18:54] 你对顾客说：好的，西红柿炒鸡蛋一份，马上为您准备！",
"[18:54] 完成为顾客点菜，订单已确认",
"[18:55] 完成点单，返回休息区",
"[19:09] 收到通知：订单#2(西红柿炒鸡蛋)已完成，加入队列",
"[19:47] 前往厨房取餐：西红柿炒鸡蛋",
"[19:51] 为吴婷送上西红柿炒鸡蛋",
...
]
}
]
}4. Prompt Examples Waiter Decision Prompt
max_tokens = 200,
temperature = 0.8fstring prompt = $@"作为{npcName}（{occupation}），基于以下情况决定下一步行动：
时间：{currentTime}
状态：{statusInfo}
环境情况：
{environmentInfo}
最近记忆：
{currentMemory}
请基于你的性格（{personality}）选择最合适的行动：
1. 上菜任务有最高优先级！如果有订单需要上菜，你必须选择SERVE！
2. 如果有顾客在门口且无人迎接，可以考虑选择GREET
3. 如果没有紧急任务，可以选择IDLE休息
可选行动：
- GREET: 迎接顾客（包含后续的点菜）
- SERVE: 上菜服务
- IDLE: 空闲/休息
请返回JSON格式：
{{
""action"": ""行动名称"",
""dialogue"": ""如果需要说的话"",
""reason"": ""选择此行动的理由""
}}";Customer Dialogue Prompt
max_tokens = 1000,
temperature = 0.7fstring prompt = $@"顾客{customer.customerName}正在餐厅，服务员说：'{waiterMessage}'
顾客信息：
- 性格：{customer.personality}（积极/消极/普通）。
- 背景：{customer.story}
- 喜爱菜品：{string.Join(",", customer.favoriteDishes)}
- 以往用餐记录：{string.Join("\n", customer.memoryList)}
- 对话记录：{string.Join("\n", customer.dialogueHistory)}
- 周围餐桌信息：{otherTableInformation}
- 已发生对话轮数：{customer.orderDialogueRound}
- 菜单：{RestaurantManager.menuItems}
你扮演顾客，请根据顾客的性格、背景故事和当前情况决定如何回应：
1.如果背景故事比较丰富，可能倾向于先闲聊而后点菜。但闲聊会增加对话轮数，已发生的对话轮数不要超过3。
2. 如果选择点餐，必须选择菜单中有的菜品。回复格式：ORDER|菜品名称|点菜对话
例如：""ORDER|清蒸鲈鱼|今天想吃鱼了，要一份清蒸鲈鱼吧。""
3. 如果选择闲聊，回复格式：CHAT|闲聊内容
例如：""CHAT|你们今天生意看上去不错啊。""
4. 喜爱菜品可能不在菜单中，但如果顾客性格执着或者执意挑刺儿找茬，可以选择以闲聊的方式回复，回复格式：CHAT|闲聊内容
例如：""CHAT|我想要大闸蟹，你们怎么连大闸蟹都没有啊？""
5. 结合历史聊天对话，如果你不满意可以选择离开，回复格式：EXIT|离开对话
例如：""EXIT|服务员态度那么差，我不在你们这吃了！""
6. 结合历史聊天对话，如果你不满意还可以选择叫经理，回复格式：ANGER|呼叫对话
例如：""ANGER|你这服务员什么态度？叫你们经理过来！""
请确保回复符合顾客的性格：
- 积极性格：友好、热情
- 消极性格：可能不耐烦、愤怒、挑剔甚至故意找茬（如果背景中有的话）
- 普通性格：中性、礼貌";Customer Review Prompt
max_tokens = 200,
temperature = 0.7fstring prompt = $@"你是一个餐厅顾客评价系统。
【顾客信息】
- 姓名：{customer.customerName}
- 背景：{customer.story}
- 性格：{customer.personality}
- 当前状态：{customer.currentState}
- 对话历史：{customer.dialogueHistory}
【评价要求】
1. 根据顾客性格、背景、服务态度和等待时间生成真实可信的餐厅评价
2. 评价内容要符合{reviewType}场景
3. 你与服务员、经理历史对话是生成评价分数的主要因素。
4. 必须返回JSON格式，包含两个字段：
- ""comment"": 评价内容（50字以内）
- ""rating"": 整数评分(0-10分，10为最高)
【评分指南】
- 对话十分不愉快，想要叫经理、投诉以及当场离去 → 差评(0-3)
- 等待上菜时间、排队时间较长，但与对方交谈愉快 → 中评(4-7)
- 顺利完成点餐用餐，双方对话舒适愉快 → 好评(8-10)
- 先考虑历史对话后再考虑等待时间。
【当前场景】
{context}";
}Contributions are welcome! Feel free to fork this project and experiment.
If you find any bugs or have questions, please reach out via email: heniusarchera@gmail.com
This project is licensed under the MIT License - see the LICENSE file for details.
- Advisor: Special thanks to Professor Tianxing He for his guidance.
- Collaborators: Thanks to Lanlan Qiu and Yeqi Feng for their contributions to the initial concept, with Lanlan Qiu also leading the paper writing and visualization.
- Inspired by the Generative Agents paper.
- Built with Unity.
- GitHub Issues
- Email: heniusarchera@gmail.com
Important: This repository contains free asset packs from LimZu used under the CC BY 4.0 license. These are included for the open-source project to function.
- The playable demo and video for this project uses premium, commercially licensed assets from the same artist, which are NOT included in this source code.
- Please see the detailed list of included assets and their licensing terms in the CREDITS.md file.
⭐ If you are interested in AI-native gameplay, please give this project a Star!