✦ Overview
David Ondrej walks through Hermes Agent's Mixture of Agents (MOA) feature — a technique that consults multiple AI models (GPT-5.5, Opus 4.8, GLM 5.2, Kimi K2.7) in parallel, then funnels their responses into a single aggregator model to produce answers that surpass what any individual model can achieve alone. The video covers the full pipeline: conceptual explanation, VPS provisioning, Hermes installation via Claude Code, MOA preset configuration with custom temperatures, a live demo building and deploying a 3D Flappy Bird game, and a passionate argument for open-source model adoption.
1 What is Mixture of Agents?
Mixture of Agents (MOA) is a new Hermes Agent feature that consults several different AI models — Grok, GPT, Gemini, Opus, or whichever models you configure — and then sends all of their responses to a single aggregator, typically the most powerful model available. Only after the aggregator synthesizes the best elements from each response does it take action and send the final answer.
The core promise: by mixing multiple agents from different providers and companies, you extract the best capabilities of each model. David claims the result surpasses even GPT-5.5 and Opus 4.8 individually — frontier-level intelligence without needing access to unreleased models like Fable or GPT-5.6.
Critically, in Hermes, the MOA preset shows up as a model — tool calling works, memory works, session context works. It behaves exactly as if you swapped to a smarter, more powerful single model.
2 The Closed Model Problem
David frames MOA as a direct response to a tectonic shift in the AI industry: companies have stopped releasing their best models. For the past five years, every company raced to release their most powerful model to claim the top spot. Now, both OpenAI and Anthropic hold back their most capable models from public access.
This means the open-source community must "take matters into our own hands" — finding ways to get smarter AI and better results without relying on these closed companies to share their best work. MOA is exactly this kind of approach: cleverly orchestrating multiple publicly available models to achieve aggregate intelligence that exceeds any single model.
3 When to Use MOA & Presets
MOA is not for simple questions like "what's the date?" or "how many calories did I consume?" — that's absolute overkill. Instead, use it for the difficult work where you don't mind spending more money for the absolute best answers:
- Hard debugging — complex multi-file issues across backends
- Code review — architectural decisions and security analysis
- Codebase architecture planning — system design at scale
- Security hardening — vulnerability analysis
Multiple Presets
You can have multiple different MOA presets tailored to specific tasks. One preset for adding new features, another for code review. Different models excel at different things — Grok is great for Twitter research but not ideal for serious coding; Opus shines at front-end work; GLM 5.2 and GPT-5.5 are strong on backend debugging.
4 MOA vs Mixture of Experts
David makes an important clarification between two similar-sounding terms that are completely different:
- Mixture of Experts (MoE) — an internal model architecture. Instead of one monolithic model, you have multiple "experts" (e.g., 8 specialists: one for math, one for physics, one for coding). Only certain parameters are active at a time, allowing a large model to run on smaller hardware. This is how models like DeepSeek work internally.
- Mixture of Agents (MOA) — an external orchestration pattern. Several complete, separate AI models each answer your prompt independently, and then a single powerful aggregator model combines their responses into a final answer. This is what Hermes implements.
MoE is about efficient inference within a single model. MOA is about combining the strengths of multiple separate models.
5 Trade-offs & Caveats
David highlights several practical realities that "most influencers will skip over":
- Cost & latency — more models means more tool calls, more tokens, higher bills, and longer response times. This is not for minimizing costs or doing quick work. It's for squeezing the absolute best out of publicly available AI.
- Prompt caching stays intact — all of the prompt caching savings you normally get still apply with MOA.
- Reference models run in parallel — the reference agents don't debate each other. They run completely independently and simply submit their responses to the aggregator. There's no back-and-forth negotiation.
6 VPS Setup
David recommends running Hermes Agent on a VPS (Virtual Private Server) rather than your local machine. The advantages:
- Dedicated computer for the agent to operate on
- 24/7 uptime — runs automations, cron jobs, and workflows even when your personal computer is off
- Run multiple agents — Hermes, Pi, OpenClaw, Codex, all on a single server
He demonstrates the setup using Hostinger's VPS with a KVM2 plan, which he says is "more than enough" for running multiple agents. The VPS comes with a dedicated Hermes Agent preset that simplifies the initial setup. He walks through choosing a server location, setting a root password, and accessing SSH credentials from the Hostinger panel.
7 Agent-Driven DevOps with CMAX
Instead of manually SSH-ing into the VPS and running commands himself, David uses Claude Code as an orchestrator agent to set everything up. This is a paradigm shift he emphasizes heavily: use an agent you already have to set up a new agent.
The CMAX Workflow
David uses CMAX (an open-source terminal multiplexer) to manage multiple panes — Claude Code in one pane, the SSH session to the VPS in another. Claude Code can control other CMAX panes, SSH into the VPS, run terminal commands, and monitor outputs.
- He creates a project directory, stores the VPS IP in a
.envfile - Tells Claude Code to SSH into the VPS (he only fills in the root password manually)
- Then instructs: "Browse the web to learn how to set up Hermes Agent on an Ubuntu VPS, then install it root level on my VPS. Do all the steps yourself."
- Claude Code completes the entire installation in 5 minutes 29 seconds
8 Configuring Hermes with OpenRouter
With Hermes installed on the VPS, the next step is powering it with an AI provider. David chooses OpenRouter for its versatility:
- Multiple providers per model — OpenRouter has nearly 20 different providers for GLM alone, with at least one active at all times
- Easy API key management — he sets a spending limit on the key (important since MOA is expensive)
- Alternative: Vercel AI Gateway — described as "very underrated," especially for GLM 5.2
Setup Steps
David runs hermes setup (full setup, not quick setup), selects OpenRouter as the provider, pastes in the API key, and chooses GLM 5.2 as the default model — which he calls "probably the best open-source model right now" making Hermes "probably the strongest open-source agent right now."
A quick test prompt ("Who are you?") confirms Hermes is running on the VPS.
9 Configuring the MOA Preset
David configures his MOA preset by telling Claude Code to set up the following reference models (all via OpenRouter):
- GLM 5.2 — open-source, strong backend coding
- GPT-5.5 — strong general reasoning and backend
- Kimi K2.7 Code — specialized coding model
- Opus 4.8 — strong front-end and creative work
The aggregator is Opus 4.8 — the model that receives all reference responses and makes the final decision.
Temperature Tuning
David adjusts temperatures strategically:
- Reference temperature: 0.9 — the reference models should be more creative, more random, exploring different solution spaces ("they're like brainstorming")
- Aggregator temperature: 0.2 — the aggregator should be consistent and predictable, making reliable decisions about which reference's approach is best
10 MOA in Action — Live Demo
With the preset configured, David launches Hermes and sends a test prompt. The output shows Mixture of Agents in real time:
- Reference 1/4 — GLM 5.2 responds
- Reference 2/4 — GPT-5.5 responds
- Reference 3/4 — Kimi K2.7 Code responds
- Reference 4/4 — Opus 4.8 responds
- Then the aggregator (Opus 4.8 at low temperature) synthesizes the final answer
David compares this to a "17th century French court" — Opus 4.8 is the king, and the reference models are advisors. Each advisor gives their opinion on different stages of the project, but ultimately the aggregator decides which advisor is right for each aspect: "For this stage, Kimi is right. For that stage, GPT-5.5 has the best opinion." It's a complete meritocracy — only the best idea wins, regardless of which model proposed it.
11 Multi-Agent Orchestration with Pi Agent
To manage the long-running MOA tasks, David uses Pi Agent (powered by GLM 5.2 Fast via Vercel AI Gateway) as a monitoring agent. Pi sits in a separate CMAX pane and:
- Monitors the Hermes process — polls periodically, sends concise status updates every 30-40 seconds
- Goes to sleep when idle — uses
sleepcommands to avoid wasting tokens, extending the interval when it detects the task will take longer - Steers when stuck — when the aggregator appeared frozen (no new tool calls for 3 minutes, stuck at 60k tokens), Pi Agent autonomously sent a steering prompt: "Status check. If npm is finished, type check any errors, keep moving toward deploy."
- Manages CMAX panes — can spin up new terminals, switch between panes, and control the workspace
The CMAX Skill
David has a dedicated CMAX skill in his skills repository that teaches agents how to use the terminal multiplexer effectively — workspace IDs, pane management, key bindings. When Pi Agent reads this skill, it knows exactly how to interact with other panes without David having to explain it each time.
12 The Open Source Imperative
David delivers a passionate argument for shifting AI spending toward open-source models. His key points:
The Broken Social Contract
Companies like OpenAI and Anthropic scraped the entire web — all of humanity's knowledge, art, code, and literature — to train their models. The implicit deal was: "They take our data and give us powerful AI in return." But now they're keeping the best models for themselves and for governments, while the public gets lower-tier offerings.
Profit Margin Myth
Many people think subscriptions like ChatGPT Pro ($200/month) are subsidized — that OpenAI gives $2,000-$4,000 of compute for $200. David argues the opposite: API profit margins are "disgusting" — likely over 90%. The subscription is a loss leader to lock users into the ecosystem. Chinese companies (DeepSeek, Kimi, GLM, MiniMax) are so much cheaper precisely because they don't have these inflated margins.
Open Weights Models
The claim that Chinese-trained models have secret backdoors is dismissed: these are open-weights models. You can spin them up on a GPU cluster anywhere — Europe, Australia, USA, Brazil — in 10-20 minutes. The training origin doesn't matter when you control the deployment.
13 Full Demo — 3D Game Build & Deploy
The practical test: David instructs Hermes (via Pi Agent) to build a full-stack 3D game and deploy it to a live URL. The result after ~20 minutes of MOA-powered computation:
- A fully functional 3D Flappy Bird game — with 3D graphics (albeit with "a bit over the top" glow effects), working game logic, spacebar controls, and hit detection
- Fully deployed on the web — live public URL, accessible in incognito mode, no manual deployment steps required
- The agent figured out deployment independently — David didn't provide any API keys, authentication, or deployment instructions. It built the app and deployed both front-end and back-end on its own.
Cost Analysis
David checks the OpenRouter dashboard after the demo:
- Initial testing: $0.20 — basic MOA test prompts
- Full 3D game build & deploy: ~$20 total
The verdict: "Not the cheapest. Definitely not the cheapest. But this is how you get frontier intelligence."
🎯 Key Takeaways
🔑 Key Takeaways
- MOA surpasses individual frontier models — by orchestrating multiple models in parallel and synthesizing via an aggregator, you get intelligence beyond any single publicly available model.
- It's a model-level abstraction — in Hermes, MOA presets appear as regular models. Tool calling, memory, sessions — everything works transparently.
- Reserve MOA for hard problems — debugging, architecture, code review, security. For simple tasks, it's expensive overkill.
- Temperature strategy matters — high temperature (0.9) for reference models (creative exploration) and low temperature (0.2) for the aggregator (reliable synthesis).
- Multiple presets for different domains — tailor reference model selection to the task type (coding, research, creative work).
- Use agents to set up agents — Claude Code or Pi Agent can install Hermes, configure it, and manage it on a VPS without manual DevOps.
- CMAX + skills = autonomous monitoring — a monitoring agent with proper skills can poll, sleep, and steer a working agent without human intervention.
- Open source models have caught up — GLM 5.2 and Kimi K2.7 compete with closed models at a fraction of the cost. Two of the four MOA references are open source.
- Closed model pricing has inflated margins — API profit margins may exceed 90%, making the "subsidized subscription" narrative misleading.
- MOA costs are real — a full build-and-deploy task cost ~$20 through OpenRouter, compared to pennies for a single-model prompt.
- Prompt caching still applies — MOA doesn't break existing cost optimizations like prompt caching.
- Self-hosting is strategic — as AI becomes essential infrastructure, depending on closed providers means ceding control over your most important tools.
🔗 Resources & Links
- David Ondrej — MOA Prompts, Configs & Presets — free resource with all prompts and presets from this video
- OpenRouter — multi-provider AI gateway used for all MOA reference models
- Hermes Agent by Nous Research — the open-source agent framework with MOA support
- CMAX — open-source terminal multiplexer for managing multiple agent panes