What happened on the two RTX Spark laptops
His first Hermes session ran on a Surface Laptop Ultra with 32GB of memory. Hachman notes that this capacity is shared between system memory and GPU memory. He reasons that less VRAM means smaller local models and possibly slower performance. He gives no measured allocation or speed figures.
On that machine, Hermes' defaults leaned heavily toward caution. It asked him to approve each mouse movement and click. He wanted a way to track tokens produced, and said he felt he was being trained to hand Hermes total control. He didn't get far before a colleague borrowed the laptop and he wiped it.
He then moved to a second RTX Spark laptop with 64GB of RAM and installed Hermes from the Microsoft Store again. This time Hermes acted autonomously without per-click approval. Hachman doesn't recall enabling that. His account doesn't explain the difference, and it could come from a setting, install state, version or something else.
Autonomy controls: check them first
A small control in the lower-right corner of the Hermes interface sets how much autonomy the agent has. According to Hachman:
- Manual requires human approval in almost every case.
- Smart handles most actions on its own.
- A third option removes human oversight entirely. He says he wouldn't enable it.
Check this setting before you hand Hermes anything on a PC you care about. That is my advice, not a documented Hermes procedure.
The cloud backend surprise
The bigger time sink was the model backend. Hachman says Hermes on the second laptop was initially set up to call a Nous cloud model that could act locally. He says this was never clearly disclosed. His cloud credits had expired and his API key was rejected, so Hermes had no model to use and sat idle.
He then downloaded a 27-billion-parameter Qwen 3.8 model to run locally. NVIDIA's RTX Spark setup guide describes the same route. It says Hermes walks users through a chat to connect apps and choose a local or cloud model. It also says to select Qwen3.8 27B under LOCAL in the Hermes prompt box to use a local model.
NVIDIA's guide gives a hardware recommendation too. For a good agent experience with local models, it recommends a model of around 30B in size, which typically requires an RTX Spark device with 48GB or more. That is relevant to Hachman's 32GB Surface, though his article doesn't test that threshold.
The practical lesson is to confirm which backend is active before judging local performance. Check whether it needs a key or credits and whether the model you want is actually installed. A "local agent" that quietly depends on a cloud model is a different thing from one running fully on your hardware. This is my inference from his experience, not a Hermes checklist.
The token-counter saga
Hachman's goals were modest: a counter for tokens produced and a tokens-per-second indicator. Hermes decided it needed a plugin from the Nous site and tried to install it, producing repeated errors. It kept retrying until he asked what was wrong. Hermes then said it needed Git. Installing Git took a few quick command-line taps.
Even after the plugin loaded and a reboot or two, the tracker didn't seem to work. Splitting the work across two sessions didn't help, because the available tokens were divided between them. Hachman says the whole detour took far longer than he had planned.
File access and other cloud options
By default, Hermes had no file access. Hachman says that granting it required adding a special MCP, and that Hermes itself advised against running it. The article doesn't name the MCP or validate the warning. Treat file access as a separate permission decision, and don't assume a security flaw from this one passage.
He also tried other cloud providers as a backend. His work ChatGPT subscription wouldn't give up an API key, and his personal Gemini account didn't seem to work. Free Gemini did work. He found it much faster, but it removed the point of going local. These are observations about his accounts, not statements of current API rules.
Vendor promises vs. one reporter's day
NVIDIA's September 3 blog promised easier local agents. It said Hermes would detect the NVIDIA GPU, select an appropriate model and configuration, and run it through integrated llama.cpp, eliminating manual model downloads and tuning. It also said one-click local model setup was available on Windows at that time. NVIDIA's October launch announcement says guided setup and easy installers help users start agents including Hermes and OpenClaw without a command line.
Hachman's day included command-line work and a manual model download. These accounts can both be true. Setup may differ by install path, app version and OOBE choices. One more detail from NVIDIA's own Hermes playbook for DGX Spark: its Quick Setup signs in through the Nous portal and doesn't offer provider selection. That page covers DGX Spark rather than the Windows Store app, but it shows that different setup routes can land on different backends. I can't confirm that this explains what Hachman saw.
Why memory matters on these machines
Microsoft's Windows Experience Blog describes RTX Spark as supporting up to 128GB of unified memory. It says Windows now allows a higher limit on total system memory accessible by the GPU, which helps load larger local AI models. It also says Hermes Agent and OpenClaw will integrate NVIDIA OpenShell and new Windows security and containment primitives. That is platform context and doesn't show those protections were active in Hachman's test.
What the piece does and doesn't show
Supported:
- Local agent setup on these two machines involved confusing defaults, a hidden cloud dependency, plugin and permission hurdles, and a lot of lost time.
- The 32GB and 64GB configurations may behave differently, though the article gives no data.
Not supported:
- That Hermes is generally broken.
- That local AI is slower than cloud AI across the board.
- That RTX Spark hardware caused the failures.
Hachman himself separates agentic tools from local AI generally. He finds cloud chatbots and local models run through apps like Ollama efficient, and he finds building his own AI apps productive. He isn't writing off local AI. He thinks local agents may work best in the background while you work on another screen, but concludes Hermes is "a bit out of reach for the average user" right now.
Quick checklist before you try this
- Confirm whether the agent is using a local or cloud model, and whether credits or keys are required.
- Make sure a local model is actually installed and selected. On a 32GB machine, expect smaller models than on 48GB or more.
- Set the approval level deliberately. Don't enable no-oversight mode casually.
- Treat file access and added tools as separate permission decisions.
- Expect to install dependencies such as Git if the agent asks for them.
The idea of an always-on local assistant is appealing, and vendors say it is getting easier. For now, Hachman's day is a reminder to budget setup time and read the permissions carefully.