NVIDIA Put Hermes and Claude on an RTX Spark PC. Then the Agent Designed a House

#nvidia #ai #machinelearning #programming

I expected the usual launch rhythm when I watched NVIDIA's RTX Spark announcement: a new chip, a wall of specifications, a few game clips, and a promise that this changes everything.

Then an AI agent started designing a house.

It took a site, concept sketches, a mood board, and written requirements. It opened Rhino, modeled the terrain and building envelope, generated an interior layout, moved the project into Blender, and used FLUX to help produce photorealistic renders. NVIDIA said the computer was running an open-shell sandbox with a Hermes harness connected to Claude Sonnet in the cloud.

That two-minute demonstration told me more about NVIDIA's plan than the hardware reveal did. RTX Spark is being built for a computer where you give an objective to an agent and the agent operates the applications.

NVIDIA's official RTX Spark product image. Source: NVIDIA RTX Spark.

Watch the house-design agent

The useful part begins at 7:29 in NVIDIA's keynote video. This embedded clip is set to play the house-design workflow through its conclusion.

Watch NVIDIA's RTX Spark house-design agent demo (starts at 7:29)

NVIDIA also posted a short behind-the-scenes video from GTC Taipei that includes the RTX Spark launch:

Watch NVIDIA's GTC Taipei behind-the-scenes video on X

What the agent actually did

The demonstration starts with inputs an architect might already have: a site, rough sketches, visual references, and a text description of the requirements. The agent then uses the applications installed on the laptop.

It opens Rhino and models the site.
It shapes the terrain, setbacks, and building envelope.
It proposes forms based on cost, comfort, and quality.
It generates walls, circulation, rooms, doors, windows, and structural elements.
It detects at least some of its own mistakes.
It exports the approved model from Rhino to Blender while preserving design context.
It renders the house and uses FLUX to make the images photorealistic.

The human does not disappear. The presenter says she can jump in, approve decisions, adjust materials, and choose the final shots. Still, the division of labor is different from the workflow most of us know. The user is directing the job rather than clicking through every step.

I have written before that developers are starting to direct agents instead of writing every line themselves. This demo applies the same idea to professional desktop software.

Claude was in the cloud

There is an important detail in NVIDIA's narration: the Hermes harness was running on the RTX Spark system, but it was connected to Claude Sonnet in the cloud.

So this was not a demonstration of Claude running completely offline on the laptop. The local computer hosted the agent environment and professional applications. A cloud model supplied at least part of the reasoning. That makes the demo less magical, but more believable.

Hybrid agents may be the practical design for a while. A local system can hold files, operate software, run smaller models, and keep long-lived processes available. A cloud model can step in when the task needs stronger reasoning. The system can also switch models as costs, privacy requirements, and model quality change.

This is why I find the harness more interesting than the logo on the model. A good harness manages tools, files, permissions, context, and the loop between an instruction and a result. If that layer works well, Claude can be replaced by another cloud model or a capable local model later. NVIDIA's own video even mentions local Nemotron models alongside Claude and Codex as possible options.

Why RTX Spark exists

NVIDIA's official product page lists configurations with up to a 6,144-core Blackwell RTX GPU, a 20-core Grace CPU, one petaflop of FP4 AI performance, and 128 GB of unified memory. CUDA runs natively, and NVIDIA is positioning the machine for creation, AI development, gaming, and agents.

The unified memory is especially relevant. Agents that combine a language model, computer vision, code execution, 3D tools, and generative media can consume a lot of memory before the user even opens a normal workload. Giving the CPU and GPU access to a large shared pool makes the system more suitable for this kind of mixed work.

NVIDIA's marketing line is unusually direct: "Your PC just went from tool to teammate." The company describes agents that run tasks, generate assets, and write code while the user sets the objective.

I am still cautious about calling a computer a teammate. Software does not share responsibility when something goes wrong. But the intended interface is clear. NVIDIA does not want RTX Spark to be judged only by how quickly it renders a frame. It wants people to imagine persistent agents using that compute all day.

Applications are becoming tools for agents

The house demo only works if an agent can interact reliably with Rhino, Blender, the file system, and the image model. That is a harder problem than making a chatbot answer a question.

Later in the keynote, Jensen Huang said Adobe had re-engineered Photoshop and Premiere for RTX Spark and made them agent-friendly through an MCP server. If major desktop applications expose stable tool interfaces, agents will not need to imitate a mouse click for every action. They can call defined operations, inspect results, and continue the workflow.

That should be faster and less fragile than screen automation. It could also be safer if each tool has clear permissions and an audit trail. An agent might be allowed to create a draft in Blender but blocked from overwriting the approved production file. That kind of boundary matters once agents can work for minutes or hours without someone watching every step.

A polished demo is not proof of reliability

The video is impressive, but it is still a launch demonstration. We do not know how many attempts it took, how much of the workflow was prepared, how often the agent gets stuck, or whether it can recover from a messy project that was not designed for the presentation.

I would want answers to some boring questions before trusting this setup with paid work:

Can I see every command and tool call?
Does the agent ask before deleting, exporting, or replacing files?
Can I restore the project after a bad action?
What information leaves the PC when a cloud model is used?
How does the agent behave when Rhino, Blender, or an MCP server returns an unexpected error?
Can the same workflow succeed repeatedly, not just once on stage?

Those details will decide whether an agentic PC is useful or merely good at producing launch videos.

This is a better argument for the AI PC

I previously looked at DGX Spark and questioned whether an expensive personal AI supercomputer made sense for normal users. I also researched when Claude-level local AI might run on an ordinary PC.

RTX Spark does not settle either question. We still need real pricing, independent tests, battery results, and evidence that the agent workflows survive outside NVIDIA's controlled demo. The use case is much clearer now, though.

A large pool of unified memory seems excessive if the computer is only waiting for someone to open a browser and type into a chat box. It makes more sense when the system is expected to keep an agent running, load models, inspect visual information, operate creative tools, and move data between several applications.

For developers, architects, 3D artists, researchers, and small teams, that could be worth paying for. For everyone else, the value depends on whether useful agents become reliable enough to save real time. A machine full of expensive compute is not helpful if its owner spends the afternoon correcting autonomous mistakes.

My take

The house-design sequence is the first RTX Spark demonstration that made the "AI PC" label feel like more than a hardware marketing category to me.

It also showed why the future is unlikely to be purely local or purely cloud-based. The PC handled the environment and applications. Hermes coordinated the work. Claude provided cloud reasoning. The user remained the director. That arrangement is less dramatic than saying a laptop independently designed a house, but it is closer to something people may actually use.

If NVIDIA and Microsoft can make application tools dependable, permissioned, and easy to inspect, RTX Spark could be an early example of a different kind of personal computer. We will spend less time opening programs one by one and more time describing a finished result, reviewing the agent's work, and deciding what is allowed to happen next.

I am not ready to call that computer a teammate. But after watching it move a building from an idea to a rendered model, I understand why NVIDIA chose the word.

References

Originally published at https://blog.jenuel.dev/blog/nvidia-rtx-spark-hermes-claude-agent-designed-house

Thanks for reading! If you enjoyed this article and like this kind of content, you're always welcome to buy me a little coffee, but only if you'd like to. No pressure at all, and either way I'm truly grateful you stopped by. ☕️