Open-source agent platform Hermes Agent now supports a new preset called Mixture of Agents (MoA). MoA is treated as a virtual model provider, not a tool, so you can pick it from the model dropdown or use the /model command. There's also a /moa [prompt] shortcut for one-off calls: it temporarily switches to the default MoA preset for a single turn, then reverts.
https://twitter.com/NousResearch/status/2070610321278988385
Here's how MoA works: Your preset defines a reference model and an aggregator model. Hermes first sends a stripped-down version of the conversation — no system prompt or tool call history — to the reference model, which produces an analysis. That analysis gets appended to the end of your latest user input. Then the aggregator model, acting as the main agent, sees that analysis and generates a final response with full tool schemas and system prompts, making any tool calls it needs.
In the upcoming HermesBench benchmark, a MoA preset using Claude-Opus-4.8 as aggregator and GPT-5.5 as reference scored 82.02%. That's 6 percentage points (about 8%) higher than running Claude-Opus-4.8 alone, and 11% higher than GPT-5.5 alone. To keep prompt caching efficient, the reference model's input is simplified for stable caching; the aggregator puts the reference analysis at the very end of the prompt to avoid changing the byte prefix of the conversation history.
You can list presets with hermes moa list and add or modify them with hermes moa configure [name]. No recursive nesting allowed — the aggregator can't point to another MoA preset. If the reference model fails (e.g., expired credentials), Hermes doesn't stop; it just passes the error info to the aggregator and continues.
MoA is now available across CLI, gateway, desktop, and TUI interfaces.