Author: Ningning
I spent some time studying the current version of the AI Agent framework Eliza of AI16Z and the differences between it and the V2 version revealed by Shawn.
Before the comparison, let's first understand the basic architecture of the current version and the V2 version.
As shown in Figure 1, the current version of the Eliza framework consists of the Interface Layer, the Core Layer, and the Model Layer.
Interface Layer
Responsible for interacting with external platforms
Supports multiple platforms such as Discord, Twitter, and Telegram
Provides API interfaces for other systems to call
Core Layer
Memory system: provides retrievable storage and document processing capabilities
Behavior system: manages custom behaviors and action execution
Configuration system: manages configurations through .env and elizaConfig.yaml
Model Layer
Supports various large language models, including OpenAI, Anthropic, Llama, Grok, etc.
Provides a unified model interface abstraction
The interaction between the layers is through standard interfaces, supporting a high degree of modularity and extensibility.
On the 15th, Shawn revealed that they are developing the Eliza v2 version, and the alpha version is expected to be released in 2 to 3 weeks.
Eliza v2 will unify the message bus and simplify the client; unify the Agent wallet; adopt a registry and override model for the model system; enhance the extensible and generic core framework; update community plugins; and achieve 100% test coverage.
The estimated architecture design of the V2 version is as follows:
Let me explain the main improvements of Eliza v2 and their positions in the architecture:
Improvements in the Interface Layer: Unified message bus
The originally scattered client integration is transformed into a unified message processing system
Each platform (Discord/Twitter, etc.) becomes an adapter to the message bus
Simplifies client development and maintenance
Provides a more consistent message processing experience
Improvements in the Core Layer: Unified agent wallet
Integrates the originally scattered wallet configurations
Provides unified multi-chain wallet management
Simplifies cross-chain operations
Extensible core framework: Redesign the core interface to make it more generic
Enhances the extensibility of the framework
Improves the plugin integration mechanism
Community plugin updates: Improve the plugin system
Improve plugin management
Ensure quality with 100% test coverage
Improvements in the Model Layer: Introduce a model registry mechanism
Support dynamic registration and override of models
Categorize models into two main types: LLMs and Embeddings
More flexible model calling methods
The new version of the architecture is more modular and unified, with clearer interactions between different components, and provides a better foundation for future expansion.
Finally, from an investment perspective, Eliza v2 will bring new growth opportunities for functional extension plugins like AgentTank.
