Week 3: Research and Planning for Expert Architectures
This week was primarily focused on researching and planning the implementation of specialized agent architectures for the tv.llm module. After exploring various approaches to improve the reliability of natural language to code translation, I settled on implementing a Mixture of Experts (MoE) and Program of Experts (PoE) architecture.
The core idea was to move away from a monolithic LLM approach to a more modular system where specialized agents would handle specific aspects of the Tölvera code generation. This week involved:
- Researching existing MoE/PoE implementations and their applicability to creative coding
- Designing the agent orchestration system
- Planning the specialized agents needed for different aspects of particle system generation
For the detailed implementation of these concepts and the actual code developed, please see Week 4: MoE and PoE Implementation, where the full architecture was built out with the CodeGenerationOrchestrator and its suite of specialized agents.