How to Use Mixture of Agents in Hermes (Two Models, One Better Answer)

Mixture of Agents lets two models think about your question separately, then hands both of their takes to a third model that writes the final answer. In Hermes, all of that happens inside a single chat message. You ask once, you can read what each model proposed, and you get one response that already weighed both.

My current blend uses gpt-5.6-luna and deepseek-v4-flash as the thinkers, with gpt-5.6-terra aggregating their output into the final answer. deepseek-v4-pro works just as well in that slot. The aggregator is the seat that needs strong reasoning, because its job is to judge two competing answers, not just polish them. On quick questions it is overkill. On subject lines, positioning, and budget decisions, it beats whatever single model I would have reached for on my own. Here is exactly how to set it up, three ways to use it, and when to leave it off.

Core Concepts about Mixture of Agents in Hermes Agent.

• Mixture of Agents (MoA) sends your prompt to two or more models, then a final model writes one answer from their combined output.
• Hermes shows each model’s individual response above the final answer, so you can audit the thinking instead of trusting a black box.
• /moa runs one question through the blend, then returns you to your normal model.
• Selecting Mixture of Agents in the model menu blends every reply until you switch back.
• The blend keeps full tool access, so it can still search the web and read files you drop into the chat.
• Each blended answer costs several model calls instead of one, so it is built for hard questions, not quick ones.

Who does this apply to

This is for marketers, operators, and builders already running Hermes Agent who want stronger output on decisions that actually matter: ad copy, positioning, budget reallocation, and campaign planning. If you have ever pasted the same prompt into three different chat windows to compare the answers, MoA automates that habit and does the comparison for you.

What Mixture of Agents Actually Does

MoA answers one question with a small committee instead of a single model. Two or more models act as proposers, each writing an independent response to your prompt. A final model, the aggregator, reads those responses and writes the answer you actually see.

The pattern comes out of published research, not vendor marketing. The original Mixture-of-Agents research paper describes a layered architecture where each layer of models receives the previous layer’s outputs as auxiliary information before generating its own response. The notable finding was that models produce better work when they can see other models’ attempts, even when those attempts are weaker than what they would have written alone.

The tradeoff is the same one you feel in Hermes: better answers, slower first token.

Inside Hermes, my blend looks like this:

• Thinker 1: gpt-5.6-luna, strong at structure, framing, and marketing voice
• Thinker 2: deepseek-v4-flash, fast and cheap, good at blunt practical takes
• Aggregator: gpt-5.6-terra, which reconciles both takes and produces the final answer. deepseek-v4-pro is an equally strong pick here.

The part that makes this genuinely useful is transparency. Hermes prints each thinker’s response above the final answer, so when the blend commits to a direction you can see which model pushed it there. That is very different from asking one model to “consider multiple perspectives” and hoping that it did.

Example 1: Run One Question Through the Blend With /moa

Use /moa when you want a second opinion on a single question without changing your session. Type the command, add your question, and Hermes routes that one message through the blend before dropping you back into your normal model.

/moa Give me 5 email subject lines for a product launch, then tell me which one to send and why.

What came back was three things stacked in one message:

1. The gpt-5.6-luna take: five curiosity-driven lines, heavier on intrigue and pattern breaks.
2. The deepseek-v4-flash take: five plain-benefit lines, shorter, more direct about what the reader gets.
3. The final answer from gpt-5.6-terra: five lines that pulled the sharpest options from both, plus a single recommendation and the reasoning behind it.

The recommendation is the part a single model usually gets wrong. Asked alone, most models pick their own cleverest line. The aggregator had two competing sets to compare, so it chose the line that was clear first and interesting second, which is almost always the one that performs on a launch send.

After that message, the session went straight back to my default model. Nothing else about the conversation changed.

Example 2: Blend Every Answer for a Whole Session

For longer thinking work, turn the blend on for the entire session instead of one message. Open the model menu, pick your Mixture of Agents blend, and every reply is blended until you switch back.

I used it to research what is currently working in email marketing for AI tools. The blend kept full tool access, searched the web, and came back with three trends, each with sources attached and a note on what to test first.

Two details made the session version worth it:

• Both thinkers read the search results independently, so the answer was not anchored to whichever source loaded first.
• Where the two models disagreed on how strong a trend was, the aggregator said so instead of flattening it into false confidence.

When the research block was done, I switched back to my normal model. Leaving a blend on for routine chat is the fastest way to burn calls for no benefit. If you want to see how model choice changes a full production workflow, I broke down the GPT-5.6 blog workflow in Hermes in a separate walkthrough.

Example 3: Point the Blend at a Spreadsheet

The blend reads files, which is where it starts paying for itself. Still on the session blend, I dropped in a Google Ads export and asked for a walk-forward strategy for the following month.

Hermes read the file, and the two thinkers split in a useful way. One went hunting for waste. The other went looking for what deserved more budget. The aggregator merged both into a single plan.

What it surfaced from the raw export:

• Two ad groups holding roughly a third of the month’s spend with no conversions in three weeks
• One ad group quietly running the cheapest cost per lead in the account while capped by its own budget
• A search term pattern pulling clicks that were never going to convert

The plan that came back was concrete: cap the leaking ad groups, move that spend to the capped winner, add the bad search terms as negatives, and hold a small test budget for the next creative round. A single model usually gives you the observation or the plan. The blend handed back both, with the numbers it drew each conclusion from.

How to Build Your Own MoA Blend in Hermes

The blend is not preconfigured. You define which models think and which model writes, then save it as a named option in your model menu.

1. Open Settings.
2. Go to Model.
3. Select Mixture of Agents.
4. Add two models as your thinkers. These run in parallel on every prompt.
5. Choose the aggregator, the model that weighs both takes and writes the final answer.
6. Give the blend a name you will recognize in the model menu.
7. Save it. It now appears alongside your other models, and /moa uses it for one-off questions.

A few rules of thumb for picking the three slots:

• Thinkers: use two different families so you get genuinely different angles, and let one of them be a cheap fast model.
• Aggregator: use a strong reasoning model, because this slot has to judge which thinker is right, not just write cleanly. I run gpt-5.6-terra here, and deepseek-v4-pro is just as capable.
• Voice: the aggregator also sets the tone of the final answer, so when two reasoning models are close, take the one whose writing you prefer.

The full configuration reference lives in the official Hermes Agent documentation if you want to manage blends from config files instead of the interface.

Single Model, /moa, or a Full MoA Session

Sub-agents and MoA solve different problems. Sub-agents split work across several jobs. MoA improves the quality of one answer. If parallel delegation is what you actually need, it is easier to run a Codex subscription inside Hermes and brief several agents at once.

When to Use It and When to Skip It

Reach for the blend when the answer is worth more than the extra calls:

• Ad copy, subject lines, headlines, and landing page angles
• Positioning and messaging decisions where you want the disagreement visible
• Budget reallocation and account analysis from a real export
• Research where one model’s first source should not decide the conclusion
• Any recommendation you are about to send to a client

Skip it for everything else:

• Quick factual questions and definitions
• Formatting, renaming, and cleanup tasks
• Long back-and-forth chat where latency matters more than depth
• Anything you would accept from your default model without a second look

The fastest way to judge it is to run one real decision through /moa, then read both thinker responses before the final answer. Seeing where two models disagree about your own work is usually more useful than the answer itself.

Frequently Asked Questions

What does /moa do in Hermes?

It routes a single message through your Mixture of Agents blend. Your thinker models each answer the prompt, the aggregator writes the final response, and the session returns to your normal model afterward.

Can I use more than two models as thinkers?

Yes. Two is the practical sweet spot for cost and latency, but a blend can hold more proposers. Adding models increases both quality ceiling and expense, so add a third only when the extra perspective is genuinely different.

Does the blend still search the web and read files?

Yes. Tool access is not affected by using a blend, so it can search, read spreadsheets and documents you drop in, and work with your configured toolsets.

Which model should be the aggregator?

Use a strong reasoning model. The aggregator reads two competing answers and decides which parts are right, so it is a judgment task before it is a writing task. I use gpt-5.6-terra, and deepseek-v4-pro handles it just as well. It also sets the tone of the final answer, so among models with similar reasoning, pick the one whose writing you like.

Does Mixture of Agents cost more?

Yes, every blended reply is several model calls instead of one. Pairing one premium model with cheaper open models keeps the cost reasonable while still producing different points of view.

Can I see what each model said?

Yes, and that is the main reason to use it. Hermes shows each thinker’s response above the final answer, so you can trace which model drove the recommendation.

Is Mixture of Agents the same as Mixture of Experts?

No. Mixture of Experts routes tokens between specialized parts inside one model. Mixture of Agents runs entirely separate models against the same prompt, then has another model synthesize their answers.