Testing Mixture of Agents: Combining DeepSeek, MiniMax, and GPT-5.4 for Multi-Model AI Decisions

Short summary

An experiment with Hermes Agent's Mixture of Agents (MoA) architecture, where multiple reference models (DeepSeek V4 Flash, MiniMax M2.7) independently analyze a prompt and an aggregator model (GPT-5.4) synthesizes their responses. The author tested it by building a Kanban board app in 14 minutes from a single prompt. The key insight is that diverse model perspectives can improve final output quality over single-model approaches.

• •Mixture of Agents sends prompts to multiple reference models in parallel, then an aggregator synthesizes responses
• •Tested with DeepSeek V4 Flash and MiniMax M2.7 as advisors, GPT-5.4 as aggregator
• •Built a functional Kanban app from a single prompt in 14 minutes using the MoA approach

Generated with AI, which can make mistakes.