15.7 Summary and Outlook

15.7.1 Chapter Review

In this chapter, we completed a full AI town project - Cyber Town. This project combines the HelloAgents framework with the Godot game engine to create a vibrant virtual world. Let's review the core content we learned.

Technical Architecture Design

We adopted a separated architecture of game engine + back-end service, separating front-end rendering, back-end logic, and AI intelligence into different layers. Godot handles game graphics and player interaction, FastAPI handles API services and state management, and HelloAgents handles NPC intelligence and memory systems. This layered design allows each part to be developed and tested independently, and also provides a good foundation for future expansion.

NPC Agent System

We used HelloAgents' SimpleAgent to create independent agents for each NPC. Each NPC has its own role setting, personality traits, and memory system. Through carefully designed system prompts, we made Zhang San a rigorous Python engineer, Li Si a product manager good at communication, and Wang Wu a creative UI designer. These NPCs can not only understand player dialogue but also respond according to their role characteristics.

Memory and Affection System

We implemented a two-layer memory system: short-term memory maintains dialogue coherence, and long-term memory stores all interaction history. Through semantic retrieval in vector databases, NPCs can recall previously discussed topics. The affection system allows NPCs' attitudes toward players to change with interaction, from stranger to close friend, with different behavioral expressions at each level. These designs make NPCs appear more realistic and interesting.

Game Scene Construction

We used Godot to create a pixel-style office scene, implementing player control, NPC wandering, interaction detection, and dialogue UI. Through the modular design of the scene system, we can easily add new NPCs, new scenes, and new functions. GDScript's concise syntax makes game logic implementation intuitive and efficient.

Front-End and Back-End Communication

We used HTTP REST API to implement communication between the Godot front-end and FastAPI back-end. Through asynchronous requests and signal systems, we ensured game fluidity - even with high network latency, player experience is not affected. The API client encapsulation allows other scripts to conveniently call back-end services, and the dialogue UI implementation allows players to naturally communicate with NPCs.

The project's technology stack is shown in Figure 15.15:

Figure 15.15 Cyber Town Technology Stack

15.7.2 Extension Directions

Cyber Town is just a starting point - there are many directions for extension. These extensions can not only enhance game fun but also explore more possibilities for AI technology in games.

(1) Multiplayer Online Support

Currently, Cyber Town is a single-player game, but we can extend it to a multiplayer online game. Multiple players can simultaneously enter the same office and interact with NPCs and other players. This requires introducing WebSocket for real-time communication and databases to persist player data and NPC states. NPCs can remember interactions with different players and maintain independent affection levels for each player.

(2) Quest System

We can design a quest system for NPCs. When a player's affection with an NPC reaches a certain level, the NPC will provide special quests. For example, Zhang San might ask the player to help debug code, Li Si might ask the player to collect user feedback, and Wang Wu might ask the player to evaluate design proposals. Completing quests can earn rewards and further increase affection.

(3) NPC-to-NPC Interaction

Currently, NPCs only interact with players, but we can enable NPCs to interact with each other. Zhang San can discuss product requirements with Li Si, Li Si can discuss interface design with Wang Wu, and Wang Wu can discuss technical implementation with Zhang San. These interactions can occur automatically in the background, and players can observe dialogue between NPCs, making the entire world appear more lively.

(4) Emotion System

In addition to affection, we can add a more complex emotion system for NPCs. NPCs can have different emotional states such as happy, sad, angry, and excited, which affect NPC reply style and behavior. For example, when an NPC is in a good mood, they'll be more willing to share information; when in a bad mood, they might be rather cold.

(5) Dynamic Event System

We can design dynamic events to make the game world richer. For example, regularly hold team meetings where all NPCs and players gather to discuss project progress; or hold birthday parties celebrating an NPC's birthday; or emergency tasks requiring everyone's collaboration. These events can increase game variety and fun.

(6) Larger World

Currently, Cyber Town has only one office scene, but we can expand to a larger world. We can add different scenes like cafes, libraries, and parks, each with different NPCs and interaction methods. Players can move between different scenes and explore a broader virtual world.

(7) Personalized Learning

NPCs can learn each player's preferences and habits. For example, if a player frequently discusses Python with Zhang San, the NPC will remember the player is interested in programming and will proactively share related content in the future. If a player likes playing games at night, the NPC will remember this time habit and be more active at night.

15.7.3 Reflection and Outlook

Cyber Town demonstrates the enormous potential of AI technology in games. NPCs in traditional games are limited by preset dialogue trees and scripts, while AI NPCs can understand and generate natural language, having real conversations with players. This not only enhances game immersion but also brings new possibilities to game design.

However, AI NPCs also face some challenges. First is the cost issue - each conversation requires calling the LLM API, which incurs certain fees. For large multiplayer online games, this cost could be very high. Second is the latency issue - LLM inference takes time, and if network latency is high, players might need to wait several seconds to see NPC replies. Finally, there's the content control issue - LLM-generated content may not be fully controllable, requiring well-designed prompts and content filtering mechanisms.

Despite these challenges, the future of AI NPCs remains full of promise. As LLM technology develops, inference speed will become faster and costs will become lower. Localized small LLMs are also developing rapidly - in the future, they may be able to run directly on players' devices, requiring no network requests at all. The combination of AI technology and games will bring players unprecedented experiences.

In Part 5's graduation project chapter, we will learn how to construct general agents using single agents and multi-agents - this will be your creative time, so stay tuned!