Generative Agents for Business Simulation Game: A Technical Report

1. Introduction

In video game design, creating engaging experiences through Non-Player Characters (NPCs) has long been hampered by traditional scripted approaches like finite state machines and behavior trees [1,2], which often lead to predictable and rigid behaviors. Recently, LLM-driven agents/NPCs have demonstrated a variety of capabilities and potential to overcome this limitation, including social connections [3,4,5,6,7], collaborative work [8,9,10,11], dialogue companionship [12,13,14], strategy application [15,16,17,18], game assistance [19,20], etc.

However, despite these encouraging advances, there are still some challenges in the applications of LLM agents/NPCs in terms of replayability and fun for video games. For example, works on social simulation [3,4] may lack clear game mechanics. In works built upon mature games like minecraft [15] or starcraft [17], the agent mechanics follow the original gameplay, thus providing limited fresh game experience. Moreover, in many of new game projects, agents are still used as auxiliary roles to enhance rather than define the core gameplay. For example, they may act as dynamic information sources to support player interactions [21], or serve as a memory-enhanced framework to ensure narrative coherence during player interactions [22].

To tackle this gap, we develop MiniBistro, a Unity-based 2D business simulation game where both the waiter and customer NPCs are generative agents [3]. The player, as the owner, strives to achieve both revenue growth and reputation success for the restaurant by interacting with this dynamic system through natural language—from directing a waiter’s tasks to personally stepping in to resolve a customer dispute. We hope this game can provide a new perspective of increasing replayability and fun in the field of AI-native games.

2. Gameplay Design

The core gameplay experience of MiniBistro stems from engaging in surprising and challenging interactions with LLM NPCs. This section outlines the main gameplay loop, the logic behind these NPC agents, and how players manage a restaurant through interactions with them. The game run effectively with GPT-4 [23], producing coherent and engaging agent responses. Our system design has been fully open-sourced to encourage further exploration by the community. For details on the NPC agent architecture and typical interaction scenarios, see Figure 1. The overall game overview is presented in Figure 2.

2.1. Main Loop

The game’s main loop is divided into three phases: preparation, operation, and review. During the preparation, the player first selects a difficulty level (Easy, Normal, or Hard), which determines the daily customer traffic and the proportion of “challenging" customers (i.e., those with lower patience and more complex needs). Subsequently, the player strategically assembles the service team for the day by selecting from a pool of LLM-driven waiters with unique backstories and personalities. During the operation phase, the player must manage various dynamic situations in real-time, such as directing waiters and handling specific customer requests. The player’s problem-solving efficiency and management skills have a direct impact on customer ratings. At the end of each business day, the system generates a report summarizing revenue, customer feedback and waiter performance. This report serves as a basis for the player’s strategic adjustments in the next cycle.

2.2. Agent Design

At the core of MiniBistro are LLM-driven waiter and customer agents, whose behavioral design roughly follows [3], encompassing observation, decision-making, action, reflection, and memory. LLMs make decisions based on the agents’ internal states, such as their personalities and memories, as well as external environments. All decision-making content and dialogues between players and waiters, among waiters, and between waiters and customers are recorded in memory in chronological order. Multi-day memories are distilled into reflections/summaries. To ensure efficient dish delivery, the chef is simplified into a state machine, allowing players to focus on service management.

• Waiter Agents. Waiters are powered by LLM and a finite state machine. The state machine handles basic tasks like “greeting customers”, “fetching dishes”, “taking breaks”, along with a “moving to destination” state that synchronizes the LLM’s processing time with real-time movement. In particular, the LLM will make personalized decisions based on the waiter’s personality, memory, and real-time status such as mood and stamina.
• Customer Agents. The satisfaction of customers are the key measure of the restaurant’s success. Each customer’s predefined personality determines their patience threshold. If their waiting time, either in the queue or for their meal, exceeds this threshold, they will leave a negative review and depart. All customer feedback, positive or negative, is ultimately compiled into the daily report, providing direct strategic guidance for the player.

2.3. Business Management

As the owner of MiniBistro, the player’s primary objective is to achieve business growth and reputation success by managing the restaurant through direct natural language commands. The management experience can be summarized in three aspects: agents who possess autonomy, intervening in dynamic emergent incidents, and managing overall operational pressure.

• Work Arrangement. The player can proactively direct waiters, but the agents do not always comply. Their final actions are autonomous decisions that balance the player’s command, customer needs, and their personalities. We observe in testing that an agent’s personality significantly influence the decision: workaholic waiters exhibit high proactivity, while lazy ones tend to rest. In addition, there may be latency in the execution of instructions. Both the autonomy and latency increase the challenge of the player’s daily management.
• Incident Handling. Second, the player must keenly observe and intervene in emergent situations involving customers. For example, a customer might complain about a food issue (e.g., “Hair in food? 1-star!!"). In such moments, the player must step in promptly to placate the customer by apologizing and offering solutions (e.g., “Sorry! We can remake or refund..") to protect the restaurant’s reputation. Improper or delayed handling of these incidents will directly result in negative reviews.
• Operational Pressure. As the game’s difficulty level (Easy, Normal, or Hard) increases, so does customer traffic and the proportion of picky customers, which adds to the operational pressure. This challenges the player’s ability to manage the business under high-pressure conditions. In our tests, by narrowing the time intervals of customer agents generation, waiters became overwhelmed, resulting in service delay and a bunch of negative reviews.

3. Conclusions

We develop MiniBistro, a Unity-based 2D business simulation game featuring generative waiter and customer agents as well as dynamic restaurant management. The implementation of MiniBistro demonstrates how generative agents can enable novel gameplay experiences and deeper immersion. We hope it provides a new perspective for developing AI-native games with replayability and fun.