Mixture of Agents

Mixture of Agents works by asking the same question to several experts and combining their answers into one conclusion. It's similar to having several chefs each taste a dish and then pooling their opinions to settle on a final recipe, instead of relying on just one chef's palate.

Since every model has its own strengths and weaknesses, relying on a single model means its weak points show up directly in the answer. By blending responses from multiple models, there's a chance for one model to cover the gaps that another model misses.

The tradeoff is that running several models at once increases computation, time, and cost. So this approach is usually chosen in situations where accuracy matters more than speed or cost.

The article explains that "an MoA (Mixture of Agents) approach that blends responses from multiple models is also possible." This is mentioned as one of several options when using the Hermes agent—where you can pick a single model or run several at once—and the method itself isn't unique to the Hermes agent; it's the name of a generally used approach.