Choosing My AI Crew

Hermes has a feature called MoA (mixture of agents). The idea is that you pick a few models to be your reporters. They each take a turn at the question or task you set out. You also pick one model to be the editor. It's the editors job to read all the drafts, throw out the bad takes, keep the good ones, and file the final piece.

Claude and OpenAI will do this when they fan out subagents but you are limited to their models. I also wrote recently about my experience with Fugu by Sakana.ai. It's also a crew of AI agents, arguing among themselves before handing me an answer. It does a really good job but it is the price of frontier models.

There are reasons to use Fugu but it is also an AI crew that one rents. Since Hermes has the MoA feature, I have to ask why rent a crew when I can hire my own and I can hire free agents who already run on my own hardware. So I set out to do just that.

the crew I hired

For the reporters I chose free and cheap. OpenRouter has a Gemma 4 free version and I have Qwen 3 30B running locally on my hardware. For the editor's chair, I was curious if Qwen 3.7 Max or GLM 5.2 would perform better. So I tested them head to head on the same tasks. Then I lined all of it up against the soloists (each model working alone) and against Fugu.

I gave the crew three real jobs, a piece of writing, a judgment call, and an explainer. I then read the outputs side by side like an editor would. The questions I wanted to answer was "does the crew I hired beat any one model working alone, and is the output anywhere near Fugu."

The editor is the most important piece. This isn't surprising to me. We are asking it to make reasoning and judgement calls. While GLM and Qwen produced final answers I'd happily ship, Qwen was painfully slow and cost roughly eight times more to do the same job. Thus, GLM was the clear choice.

The crew beat the soloists exactly where I thought it would. On the judgment call, the edited answer was genuinely sharper than any single draft. It landed on a cleaner decision rule than either reporter found alone. On the easy explainer, one model already nailed it on the first pass, and the crew just tidied the edges. Basically the crew earns its keep on the hard stuff and adds nothing to the easy stuff.

My local Qwen, the reporter I expected to be the weak link, pulled its weight. And free Gemma filed the best first drafts on two of the three jobs. Again, this was for free. The entire crew came in around two-tenths of a cent per task. Fugu, doing the same shape of work, ran somewhere between 30 and 100 times that.

what I decided to use

I made the mix of GLM in the editor's chair, free Gemma and my local Qwen as the reporters as the setup in Hermes. With Hermes I can also choose to configure different sets of models for different tasks I might run. It's why I like Hermes, I get choice and control. The MoA crew is slower than one model by design. This is a tool for when the answer matters more than the millisecond.

once again, avoid vendor lock-in

I keep writing about vendor lock-in because I think people need to hear about it. Fugu sells convenience, one call, zero setup, a crew assembled for you. That's worth something. Claude and OpenAI are of course really good but if you are relying solely on them, you are taking a risk.

Sure, there is time invested and testing when building your own setup. However, I can replace any of them tomorrow, I can reduce costs, and I am the one in control.