Mixture of Agents: Combining GPT-4.5, Opus, GLM, and Kimi to Beat Any Single Frontier Model
Mixture of agents (MoA) is the core concept: four reference models — GLM 5.2, GPT-4.5, Kimi K2.7, and a second Opus instance — run in parallel, independently answering a prompt, then hand their responses to Claude Opus 4.8 acting as an aggregator/final decision-maker. David stresses this must not be confused with mixture of experts, which is a model architecture where sub-networks share compute inside a single model. MoA is multiple complete, separate models collaborating externally.
David's central argument is that the biggest structural problem in AI right now is that frontier labs have stopped releasing their best models publicly. Both Anthropic and OpenAI are believed to hold back significantly more capable systems from API and subscription access. MoA partially compensates by combining the best *available* models to approximate capability that would otherwise be inaccessible. He also argues that open-weights Chinese models — GLM 5.2 and Kimi K2.7 — have now caught up to, or surpassed, the publicly released closed models, and can be self-hosted in Europe, the US, or elsewhere on a small GPU cluster in minutes, with no inherent backdoor risk.
The practical demo shows Hermes running MoA as its default model preset inside the Cmax/tmux environment, with a secondary PI agent (powered by GLM 5.2 fast for cost efficiency) monitoring Hermes, injecting steering prompts when it detects stalls, and restarting it if it crashes. Key operational tradeoffs surfaced: - Cost and latency increase meaningfully — the 3D game build ran ~20 minutes and ~$20 - Prompt caching savings are preserved because the system prompt cache stays intact per model - Reference model temperatures should be higher (creative/exploratory); the aggregator should be lower (consistent, decisive) - MoA makes most sense for hard tasks: architecture planning, security hardening, complex debugging — not routine queries
The session closes with a broader philosophical argument: the defining skill of the AI era is asking better questions and thinking abstractly, not executing tasks manually. David urges viewers toward self-hosting and open-weights models, warning that ceding control of AI infrastructure entirely to Anthropic or OpenAI carries long-term strategic risk. The PI agent autonomously writing structured markdown prompts for Hermes — rather than the human doing so — is offered as a concrete illustration of this principle.
David demos a 'mixture of agents' setup inside his Hermes AI framework, routing prompts through four parallel reference models before an Opus 4.8 aggregator produces a final answer. The approach claims to outperform any single publicly available model, including GPT-4.5 and Claude Opus, at the cost of higher latency and spend. The session ends with the system autonomously building and deploying a 3D Flappy Bird game in under 20 minutes for roughly $20.
The defining skill of the AI era is the ability to ask better questions. It's to have better logic. It's to think more abstract rather than saying do this task, do this task.
these agents can write better prompts than me and you as well. So like if you really express yourself properly, use the agent to orchestrate other agents and you'll get better results than you doing it yourself.
Optimizing input tokens — getting the right compressed information in front of you quickly — is as important as optimizing output tokens in large-scale agentic workflows.
Using mixture-of-agents lets you access intelligence approaching GPT-5.6 or 'fable-level' capability that is not publicly available as a single model.
it took 5 minutes 29 seconds for cloud code to set this up
Mixture of agents consults several different AI models right Grog GPD Gemini opus whatever you want. Then it sends it to a single aggregator which usually is the most powerful one and only then it takes an action and sends you an answer.
it actually defeats the best available models publicly even GBD 5.5 and Opus 4.8 which currently are the top models available to us
both OpenAI andropic have very powerful AI models that are not publicly available
in Hermes, the preset shows up as a model, so it doesn't mess up anything. The tool calling works, memory works, session context works, everything works the same.
you can use your existing subscriptions to run this at no extra cost
Grock models are great at Twitter research, right? But you probably don't want to use Grog for serious coding
Elon rumored their Grock 4.5 is going to be released soon
Mixture of experts is a architecture. Instead of having like a one large model, it's different experts. So maybe you have eight different experts and you know one of them is more about math, one of them is more about physics, one of them is more about coding. that allows you to have a large model that normally would require a more powerful computer running on a smaller computer because only certain amount of parameters are active
mixture of agents feature that we're talking about. This is several separate AI models that are complete. Each answer your prompt and then it goes to a single aggregator, single powerful model that combines it to a final response.
your chrome cache stays intact. So, all of the prone caching savings are still relevant here.
all of the models that are reference models, they don't debate between each other. They run in parallel, completely independent, and they just submit their responses to the aggregator.
The KVM2 one is more than enough. Allows you to run multiple different agents on a single server. Hermes agent, Pi, Open Claw, Codeex, whatever you want to run, the KVM2 plan is enough.
GLM 5.2 is probably the best open source model right now. So Hermes agent is probably the strongest open source agent right now.
Reference one of four. We have GLM 5.2. We have reference 204 5.5 GB 5.5. Third reference agent is Kimik K2.7 code. And the fourth one is open 4.8 and it's running.
on back end debugging you know GLM 5.2 and GPD 5.5 are really good on front end open 4.8 it is better than these models
my PI agent is powered by Versalia gateway using GLM 5.2 fast which actually has much faster inference than any of the endpoints on open router
Mixture of agents is the default model inside of Hermes now.
If Hermes got stuck or crashed, PI agent would restart it automatically because the skills were set up properly.
GLM 5.2 is really fast but cannot see things because it doesn't have vision.
Right now, GLM 5.2 is the best open source model in the world.
Total spend after the PI-agent monitoring session reached $1.80 before the 3D game build began.
Two reference agents is the minimum number where mixture-of-agents makes sense.
Models like GLM 5.2 and Kimi K2.7 have caught up to cutting-edge closed models and may surpass publicly available closed models within a month or two.
Open-weights models trained in China do not inherently have a secret backdoor; they can be self-hosted on a GPU cluster in Europe, Australia, the USA, or Brazil in 10–20 minutes with four B300 GPUs.
Running PI agent on GLM 5.2 in fast mode is significantly cheaper than running the same agent on Opus 4.8.
Anthropic and OpenAI have profit margins on API usage that are probably over 90%, which is why their subscriptions can be priced lower as a loss-leader while enterprise API customers are highly profitable.
Because everything runs inside tmux (Cmax), the PI agent can access other Cmax panes, read their output, and launch new panes or workspaces autonomously via a dedicated Cmax skill.
The Hermes agent built a 3D Flappy Bird game and deployed it to a live public URL without the user providing any API keys or authentication for front-end or back-end deployment.
The full 3D game build took almost 20 minutes and cost approximately $20 in API spend.
Opus 4.8 on OpenRouter has had many outages lately, which may have caused Hermes to stall mid-run.
always put a limit here. I'm going to click create and let's see if the setup is ready. Okay, it's still setting up peration. The nice thing is I can zoom out this CMAX pane
the biggest problem in the entire AI industry right now companies have stopped releasing their best models
more models means more tool calls, more tokens. You will pay more and it's taking longer.
OpenAI and Anthropic scraped all of humanity's knowledge — art, code, literature — to train closed models, then kept those models proprietary rather than releasing them, violating an implicit social contract.
You will wish you will wish that you are self-hosting your AI and that nobody has control over your AI models. If you give this power to OpenAI or Anthropic, you're going to regret it massively.
you want to use it for the difficult stuff. Hard debugging, code review, planning the codebase architecture, security hardening. You really want to use it where you don't mind spending more money for the very best answers that AI can give you.
mixture of Asians should not be confused with mixture of experts. Okay, these are two similar sounding terms that are completely different thing.
the reference models or agents, they're like more brainstorming, right? they should be more creative, more random. But for the aggregator, we want it to be more consistent and predictable. So, we want a lower temperature.
PI agent runs sleep timers (sleep 6, sleep 10, sleep 12) to monitor the Hermes agent, pausing to avoid wasting tokens whenever Hermes hasn't finished.
Hermes mixture-of-agents uses four reference models — GLM, GPT-4.5 (GB 5.5), Kimi K2.7, and a second Opus instance at a different temperature — whose suggestions are aggregated by Opus 4.8 acting as the final decision-maker.
The mixture-of-agents aggregator acts like a king in a 17th-century French court: Opus 4.8 is the king/aggregator, and each reference model (GLM, GB 5.5, Kimi K2.7, second Opus instance) is an adviser whose opinion is weighed on merit per task stage.
the best way to do that is on a VPS. That way you give the agent an entire computer. It can act upon. It can run 24/7 running all of your automations, chrome jobs, workflows, anything you have autonomously even when your own computer is turned off.
if you use the coupon code David, you'll get additional 10% off
Go to open router.ai go to top right create a account. If you don't have it just takes 20 seconds to create it. Go to credits charge up couple dollars. You don't need $300. Just do $5 to $10.
PI agent creates a temporary markdown file at /temp herbuilt-task and passes it as a prompt to Hermes, letting the agent write better-structured prompts than a human would.
PI agent detected that Hermes was frozen at 60k tokens with no new tool calls for 3 minutes and sent a steering prompt: 'status check. If np is finished type check any errors keep moving toward deploy.'
“The defining skill of the AI era is the ability to ask better questions.”
Shared with Earmark · earmark-ai.com