Woofun AI reports that the open-source Hermes Agent platform has officially integrated the Mixture of Agents (MoA) preset as a virtual model provider. Users can now select this preset via the /model command or desktop dropdown, and utilize the /moa [prompt] syntax for quick, single-round calls that automatically revert to the previous model. The mechanism employs a reference model to generate analysis insights from simplified inputs, which are then appended to user queries for an aggregate model to process alongside full tool schemas and system prompts. This architecture optimizes Prompt Cache efficiency by maintaining byte stability in historical dialogue prefixes.

In upcoming HermesBench tests, the MoA preset combining Claude-Opus-4.8 as the aggregate model and GPT-5.5 as the reference model scored 82.02%, marking a 6-percentage-point gain over Claude-Opus-4.8 alone and an 11% improvement over GPT-5.5 alone. The system supports CLI, gateway, desktop, and TUI interfaces, allowing users to manage presets via hermes moa commands while strictly prohibiting recursive nesting. Error handling ensures task continuity by passing credential failures into the aggregate model's context rather than interrupting execution.