Technology
How does it work?
Last updated
How does it work?
Last updated
AI agents play a pivotal role in the framework. There are two types of AI agents in the system, both being LLMs with different areas of expertise (finetuned with distinct corpora and evaluated using corresponding benchmarks).
Development and Generation Agents: These include five LLM agents, each serving a unique role in the development process. They collaborate to facilitate the creation of new virtual environments, architectures, and content.
Autonomous Virtual Agents: These are autonomous AI agents introduced into the virtual world, functioning as NPCs. They interact dynamically with the virtual environment and other agents, enriching and enhancing the content and user experience within the virtual world.
In scenarios where user prompts lack clarity or specificity, Agent 1 performs prompt refinement to facilitate smoother downstream generation steps. Additionally, as English constitutes the majority of the fine-tuning corpora, most LLM agents perform best with English prompts. Agent 1 is also responsible for translating prompts into English, ensuring that the AIGG framework accommodates users from diverse linguistic and cultural backgrounds.
A medium-sized NLP model is employed to classify user tasks. This model determines whether a prompt aims to:
Create a new game environment or world.
Introduce new content, such as the addition of autonomous AI agents, into an existing environment.
Agent 2 is responsible for decomposing tasks parsed from the user prompt into manageable sub-tasks. Based on this breakdown, Agent 2 establishes the foundational architecture of the required code, which Agent 3 then elaborates upon by generating detailed implementations.
Agent 3 generates high-quality, efficient, and well-structured code based on the task breakdown provided by Agent 2. This includes crafting modules, integrating components, and generating code adheres to industry standards.
Agent 4 takes charge of simulating and testing the environment based on task classification results:
New Environment Creation: If the prompt indicates creating a new environment from scratch, Agent 4 directly simulates its implementation and tests its functionality.
Content Integration: If the prompt involves introducing new content to an existing environment, Agent 4 retrieves the implementation of the current environment, incorporates the new content’s code, and then simulates and tests the updated version.
Agent 5 analyzes the simulation and testing results. If any issues or concerns are detected, Agent 5 provides modification suggestions and communicates these to Agent 3. Agent 3 subsequently regenerates or revises the code based on the feedback.
Additionally, Agent 5 operates with a predefined repetition counter. If the number of allowed iterations is exceeded, Agent 5 escalates the process by engaging Agent 1 to refine and enhance the original prompt, ensuring clearer and more actionable inputs for subsequent iterations.
