Skip to content
vavio

Can Hermes ask several models and combine their answers, and is it worth the cost?

Yes. Hermes Agent has a built-in Mixture of Agents (MoA) mode: two or more advisor models each give their view on your message, and one acting model reads that advice and writes the answer, using Hermes’s tools as usual. It costs more than one model. The acting model is billed for every step of the job, and by default each advisor is billed once per message you send. With Hermes’s default preset, a reply that takes six steps makes eight model calls instead of six, and the advisors read your whole conversation each time.

This follows Hermes Agent v0.23.1, commit 3b42157, checked on 20 November 2026. Upstream calls the advisors reference models and the acting model the aggregator. We reviewed official documentation and source, including how Vavio sets up Hermes. We did not run a Mixture of Agents preset, judge its answers or measure its cost.

How does Mixture of Agents work?

A preset names the advisors and the acting model, each as a provider and a model. When a preset is your model, every message goes like this:

  1. The advisors run first, side by side, without tools. Each one gets a text copy of your conversation: your messages, Hermes’s replies, the tools it called and a preview of each tool result (up to its first and last 2,000 characters). They don’t get Hermes’s system prompt or its tools.
  2. Their answers are passed to the acting model as private guidance, in a block you can see in the chat.
  3. The acting model does the work. It has Hermes’s normal tools, so it can run commands, read files and browse, and its reply is the answer you see.
  4. With the default setting, the advisors don’t run again during that reply. The acting model works through the rest of the job alone, using the guidance it already has. Your next message starts the cycle again.

If one advisor fails, for example because its provider isn’t connected, its block shows the failure and Hermes carries on with the others. If every advisor fails, the acting model works alone for that message.

How do I turn it on?

A preset appears as a model under a Mixture of Agents provider, wherever you pick a model:

  • In the Hermes app, choose a preset from the MoA presets section of the model menu, or set presets up under Settings → Model → Mixture of Agents. The settings page labels the acting model “billed for the run”.
  • In a chat, type /model <preset> --provider moa to switch to a preset, or /moa <your question> to send one message through the default preset and switch straight back to your usual model.
  • From a terminal, or by asking Hermes to run it, hermes moa list shows your presets and hermes moa configure <name> creates or changes one.

Check the default preset before you use it. In v0.23.1 it has two advisors, gpt-5.5 through a ChatGPT or Codex sign-in and deepseek/deepseek-v4-pro through OpenRouter, and anthropic/claude-opus-4.8 through OpenRouter as the acting model. Unless you have connected both of those accounts, change the preset to models on providers you use. An advisor you can’t reach is simply skipped, but the acting model has to work, because it writes every reply.

How many model calls does one answer make?

One message can take several steps: each time Hermes calls a tool and reads the result, the acting model is called again. The advisors run on a schedule set by the preset’s fanout setting:

fanout Advisors run Calls for one message (2 advisors, 6 steps)
None (one model, no preset) — 6
user_turn (the default) Once, at the start of each message 2 + 6 = 8
every_n:3 At steps 1 and 4, then every third step 4 + 6 = 10
per_iteration At every step 12 + 6 = 18

In general, with A advisors and a reply of S steps, the default makes A + S calls. The six-step reply is an example to show the arithmetic, not a measured job.

Count size as well as number. Every advisor reads its copy of the conversation, up to the length its model accepts, so a long chat costs more on every advisor, not only the acting model; see why a long conversation’s bill grows. Upstream says the preset doesn’t break the acting model’s prompt cache, so switching to a preset costs the same in cache terms as any other model change.

Who bills each call?

Each model’s provider bills its own calls. The acting model handles every step, so on a long reply with many tool steps it usually carries most of the cost; upstream says nearly all. On short replies in a long chat, the advisors can be a large share, because each one reads the conversation too. If your main model is on a subscription but the preset’s acting model is on another provider, the run is billed to that other provider, not your subscription. hermes moa list and hermes moa configure warn you when that happens.

Background tasks move to the acting model too. While a preset is your main model, Hermes’s side tasks, such as compressing a long chat, naming it or reading an image, run on the preset’s acting model unless you’ve set them to a specific model under Settings → Model → Auxiliary models. With the default preset, that is Claude Opus 4.8. Check each model’s price with its provider.

The advisors’ providers also see your conversation, including those tool-result previews. Hermes’s moa.privacy_filter setting, off by default, hides email addresses, formatted phone numbers and credential-like text in what the advisors send back: in the chat and saved traces, and with full also in what the acting model reads. It doesn’t change what the advisors are sent. Use only providers you’d trust with the chat on its own.

Is it worth the cost?

Upstream’s guide publishes one benchmark, without describing the method: on HermesBench, an anthropic/claude-opus-4.8 acting model advised by openai/gpt-5.5 scored 0.82, against 0.76 for Opus alone and 0.74 for GPT-5.5 alone. That is upstream’s result for two expensive models. It says nothing about a preset built from cheap models, and we haven’t tested one.

You want Simpler choice
A better answer to one hard question or a big decision /moa <question>, then back to your usual model
Better results on most of your work One stronger model as your main model; see what changes when you switch Hermes’s model
A cheaper recurring job A cheaper model for scheduled jobs, not a preset
A big job done faster in parallel Helpers, which split the work instead of asking for opinions

Mixture of Agents fits a question where a second opinion changes the answer, such as a plan, a review or a decision with real trade-offs. For routine replies and scheduled checks, the extra calls usually buy you little. If you mostly choose models for low cost, keep presets for occasional hard questions. To stop the advisors without deleting a preset, switch it off (enabled: false), and the acting model runs alone like a normal model.

Does it work on a Vavio computer?

From the setup source, yes, but we haven’t run a preset on one. Vavio doesn’t change any Mixture of Agents setting, so the defaults above apply, including the default preset that needs both accounts. The Vavio chat is Hermes’s own app, where upstream documents the model menu and the Settings → Model → Mixture of Agents page. Vavio keeps the computer and Hermes running, so a long preset reply carries on after you close your laptop.

Vavio provides a managed Hermes computer for $14.99 per month, with model usage billed separately by your providers. A preset adds model usage, not hosting cost. Compare the current hosting offer, or read how to use Hermes without an OpenAI or Anthropic account.

Sources

Written by Vavio with AI assistance and checked against pinned official Hermes v0.23.1 documentation and source, and against Vavio's own computer setup source. This is a source review, not a quality test or cost measurement.

More from the Vavio blog: Hermes Agent guides