Hermes Agent Launches MoA Feature: Multi-Model Collaboration Outperforms Single Model

According to the Beat Dynamics Monitoring, the open-source intelligent agent platform Hermes Agent officially supports the Mixture of Agents (MoA) preset. MoA has now transitioned to a virtual model provider and is no longer listed as a traditional underlying tool. The agent platform presents the MoA preset as a virtual model, allowing users to directly select it through the /model command or the model dropdown menu in the desktop application. In addition to model switching, users can also achieve one-time quick calls using the /moa [prompt] syntax. The system temporarily enables the default MoA preset in a single dialogue round, automatically reverting to the previous model after completing the response.

In the MoA operational mechanism, the user's selected preset contains a reference model and an aggregate model. The system first calls the reference model, inputting a simplified dialogue text without system prompt words or tool invocation history, allowing the reference model to generate analysis insights. The analysis insights from the reference model are then appended to the end of the latest round of user input. The aggregate model, acting as the agent, receives the analysis insights, generates the final reply under the full tool schema and system prompt words, and executes the tool invocation.

In the upcoming HermesBench benchmark test, the MoA preset using Claude-Opus-4.8 as the aggregate model and GPT-5.5 as the reference model achieved a score of 82.02%. Compared to running Claude-Opus-4.8 alone, this represented a 6-percentage-point increase (approximately 8%), and compared to running GPT-5.5 alone, there was an 11% improvement. To ensure the efficiency of the Prompt Cache, the dialogue input for the reference model is simplified to maintain a stable cache; the aggregate model places the reference insights at the end of the prompt, ensuring the byte stability of the historical dialogue prefix to avoid disrupting the cached context.

Users can view presets by running hermes moa list, or add and modify presets using hermes moa configure [name]. In preset configuration, the system strictly prohibits recursive nesting, where the aggregate model of an MoA preset cannot point to another MoA preset. When the reference model encounters errors such as credential invalidation, the Hermes Agent does not interrupt the task but instead passes the error information into the context of the aggregate model and continues execution. Currently, the MoA mechanism is synchronously enabled in the CLI, gateway, desktop, and TUI interfaces.