Hermes Agent Launches Mixture of Agents Feature, Benchmark Score Hits 82.02%

According to Beating, Hermes Agent has formally enabled Mixture of Agents (MoA) presets on its open-source agent platform. The MoA feature now appears as a virtual model rather than a tool, accessible via the /model command or desktop app dropdown, with quick invocation available through /moa [prompt] syntax.

In HermesBench testing, an MoA preset using Claude-Opus-4.8 as the aggregation model and GPT-5.5 as the reference model scored 82.02%, a 6 percentage point gain over Claude-Opus-4.8 alone and an 11% improvement versus GPT-5.5. The system architecture separates reference models, which generate analysis opinions from simplified input, from aggregation models, which receive those opinions and produce final responses with full tool execution. MoA is now live across CLI, web gateway, desktop, and TUI interfaces.