Descriptions:
This tutorial walks through setting up Hermes Agent’s new Mixture of Agents (MoA) feature — a capability that routes prompts through multiple AI models simultaneously before passing their outputs to a single aggregator model for a final synthesized response. The video opens with an important conceptual distinction: MoA is not the same as Mixture of Experts (MoE), which is a model architecture. MoA is an orchestration pattern that coordinates entirely separate models from different providers, and the presenter argues that this approach can outperform any single publicly available model, including GPT-5.5 and Claude Opus 4.8.
The practical setup uses Claude Code as an automated manager to configure Hermes running on a VPS, with the specific reference models being GLM 5.2, GPT-5.5, Kimi 2.7 Code, and Opus 4.8 — all accessed via OpenRouter — and Opus 4.8 serving as the aggregator. A key selling point highlighted is that MoA presets appear inside Hermes as if they were a single model swap, meaning tool calling, memory, and session context all continue working normally. The presenter also explains how to build multiple specialized presets for different use cases, such as one optimized for code review and another for feature development.
The video is honest about tradeoffs: MoA consumes more tokens, costs more money, and takes longer than single-model calls, making it best suited for high-stakes tasks like architecture planning, security hardening, and complex debugging rather than simple queries. All configuration prompts and presets shown in the video are made available as free downloads linked in the description.
📺 Source: David Ondrej · Published June 29, 2026
🏷️ Format: Tutorial Demo
