Open-source agent platform Hermes Agent now supports a new preset called Mixture of Agents, or MoA.
https://twitter.com/NousResearch/status/2070610321278988385
MoA has been rebranded as a virtual model provider, rather than a traditional tool. Users can select it via the /model command or the model dropdown in the desktop app. Beyond switching with /model, you can also use the /moa [prompt] syntax for a one-off quick call. The system temporarily engages the default MoA preset for a single conversation turn and then automatically reverts to the previous model.
In how MoA works, the preset you choose contains a reference model and an aggregation model. The system first calls the reference model, feeding it a simplified conversation text without system prompts or tool call history, so it generates analysis. That analysis is then appended to the end of the latest user input. The aggregation model — the primary actor — receives that analysis and produces the final answer with the full tool schema and system prompts, executing tool calls as needed.
In the upcoming HermesBench benchmark, an MoA preset using Claude-Opus-4.8 as the aggregation model and GPT-5.5 as the reference model scored 82.02%. That's 6 percentage points (about 8%) better than running Claude-Opus-4.8 alone, and 11% better than GPT-5.5 alone. To keep prompt cache efficient, the reference model's dialogue input is simplified for stable caching; the aggregation model places the reference opinion at the very end of the prompt, ensuring byte-level stability of the conversation prefix so cached context isn't broken.
Users can list presets with hermes moa list or add and modify presets with hermes moa configure [name]. Recursive nesting is strictly forbidden — the aggregation model of an MoA preset cannot point to another MoA preset. If the reference model encounters an error like expired credentials, Hermes Agent doesn't abort; instead, it passes that error info into the aggregation model's context and continues.
MoA is now live across CLI, gateway, desktop, and TUI interfaces.