The five parts of an AI agent, and which ones to own

On October 21, 2026, at NVIDIA GTC Berlin, Jensen Huang spent an hour and a half on AI factories, Vera Rubin racks and gigawatts. Then, near the end, a little robot duck learned to walk. And the voice in the demo said: "I use Hermes as my agent harness."

I laughed out loud. Hermes is what I've been running my week on for months. Locally, on my own workstation, with a full setup around it. Seeing it named on the GTC stage felt a bit like hearing your favourite small band on the radio.

But the duck is not the interesting part. The interesting part is how NVIDIA described an agent: five parts, each one something you can own or rent. That breakdown works just as well for a company of 50 people as for one of 50,000. So let's go through the five, and then look at which ones a small company should actually own.

Watch the full keynote on YouTube (NVIDIA GTC Berlin, October 21, 2026, 1:33:57). The five parts of an agent start at 43:16, and the Hermes demo is at 1:22:24.

What he actually said

At 43:16 he breaks it down: "Models, sandbox, harness, knowledge and tools. These are the five elements of an agent."

Short version:

• Model: the thing that thinks and writes. Closed or open.
• Harness: the layer that runs the agent. Memory, skills, scheduling. At 48:57 he lists "OpenClaw, Hermes" as harnesses you can use.
• Sandbox: the box the agent runs inside, so it can't touch everything.
• Knowledge: your documents, your data, how your company works.
• Tools: the software the agent is allowed to use.

Then came the line I liked most, at 44:36: "Please don't touch my SSH keys. That's the please don't approach. But hope isn't security."

That's exactly where a lot of AI setups are. We write nice instructions into the prompt and hope the model listens. His answer is to put the rules outside the agent. In the keynote's words: "Enforcement happens outside the agent, not just in LLM guardrails" (45:19), and "A compromised agent can't leak a key it never had" (45:50).

On models, he was refreshingly unromantic: "Every company in the future will use both closed models as well as open models that you customize" (50:10). Generic work goes to the big general models. Work that "captures too much of your proprietary knowledge" (50:02) is where your own model starts to make sense.

And then at 1:06:33, the sentence I'd put on a fridge: "the knowledge and tools. And those knowledge and tools belong to you."

How that maps to my own setup

I didn't build my setup from a keynote. I built it by trial and error, and some of it was not my proudest work. But when I lay his five parts over what I actually run, they fit almost one to one.

• Harness: Hermes. It runs on my Fedora workstation in a container, with its memory self-hosted on a small Hetzner server in Germany. It holds my skills, my scheduled jobs, my context. This is the part I own.
• Models: a mix, and it changes. Frontier models for the hard thinking, cheaper models through Ollama Cloud for the everyday stuff, local models when I want something to stay on my machine. The demo in Berlin did the same thing: run locally, and "burst larger workloads to cloud" (1:22:37) only when needed.
• Sandbox and rules: this is the part I took most seriously. The steps that really matter, like publishing, sending messages to people, money and production changes, wait for my approval. Not because I asked the agent nicely. Because the system is built so it can't do them alone.
• Knowledge: my notes, my writing, my procedures. Stored in files and systems I can open, read and move.
• Tools: the boring ones. My CMS, my calendar, my repos.

Hermes is my main harness and where my full setup lives. Alongside it, I also use Grok Bot, a hosted assistant that gives me the same kind of experience I love in Hermes. It helps with research, drafts, checking my channels and preparing things for my approval. I don't own its insides, and that's fine. I own the knowledge it works from and the gate it has to pass.

That's the split Jensen described, just smaller. Own the harness, the knowledge and the rules. Rent the models and the extra tools where that's the smarter deal.

What a 20-200 person company can realistically own

Here's the honest part. Most companies with 50 people won't build an AI factory. They don't need NVIDIA's agent toolkit to get started either. But the five parts still help, because they let you decide which pieces you care about owning and which you're happy to rent.

My rough take for a company that size:

• Rent the models. They change every few months. Staying flexible here beats betting on one.
• Rent most of the tools. You already pay for them. The question is just which ones an agent may use.
• Own your knowledge. Price lists, procedures, the way you handle a complaint. If that only lives inside one vendor's chat history, it isn't really yours.
• Own the rules. What an agent may touch and when a person takes over. That fits on one page, and it should be your page.
• Harness and sandbox: this one depends. For many teams a good hosted product is fine at the start. If AI becomes central to how you work, owning this layer gets more important, because that's where your memory and habits pile up.

You don't need all of this at once. Honestly, trying to own all five at once is how projects stall.

Where I'd start

Don't start with all five parts. Pick one piece of recurring work and ask which of the five you'd want to own for it.

Maybe it's the weekly quote follow-up, or the onboarding checklist, or answering the same ten customer questions. For that one thing, the answer is usually clear pretty fast. You'll want to own the knowledge. You'll want to own the rule for when a person steps in. The model can be whatever works best at the time.

If you've got that one piece of work in mind, the free fit check is where we look at it with you. No system access needed.

I'm not affiliated with NVIDIA or Nous Research. Quotes are from the keynote video, lightly cleaned up from the auto-captions.