A company laptop for AI agents: what to buy in 2026

The finance director has read that the new laptops “run AI agents locally” and wants the next batch to be ready for it. The refresh is due in the spring anyway: thirty machines, three or four years of service each. The question that lands on IT is simple to ask and easy to get wrong: should the agent live on the laptop, or somewhere else?

It sounds like a hardware question. It is really a question about where a piece of work runs when the person who started it closes the lid and goes home.

What has changed this month

On 16 October 2026 the first RTX Spark laptops went on sale. According to NVIDIA they run Windows 11 on Arm on a single superchip, a Blackwell RTX GPU with a Grace CPU built with MediaTek, with up to 128 GB of unified memory. Models come from ASUS, Dell, HP, Lenovo, Microsoft and MSI, from USD 2,599 depending on the region. NVIDIA ships installers for local agents and a model tuned for tool calls, and says agent permissions are handled by its OpenShell secure runtime together with Microsoft Execution Containers. Its own framing is that the PC changes “from a tool into an everyday teammate”.

Five days later NVIDIA published a reference design called Secure Agent Workspace that takes the opposite view of the device. In it, the user’s laptop is only the presentation layer. The agent runs in a managed remote environment that keeps going whether or not anyone is connected: the workspace is not the session.

Both are vendor descriptions, and both are worth reading as such. Together they mark out the choice you now have.

The real cause: an agent is a job, not an app

A chat assistant works while you type. An agent works for hours on your behalf, reads mail, files, calls tools, and may still be busy when you leave. That changes what the laptop is for. If the agent lives on the laptop, the laptop has to be awake, online and with its owner for the job to finish. If it lives elsewhere, the laptop is a window, and the thing that needs managing is the place behind the window.

Where the laptop decision goes wrong

Buying the spec, not the workload. 128 GB of memory is what a large local model needs. Most of the thirty people in the scenario will use agents for mail, documents and the CRM, and will never load one. A fleet bought for its most demanding user pays that premium thirty times.

Forgetting the closed lid. A local agent stops when the laptop sleeps, loses Wi-Fi in a train, or goes into a bag. For a ten-minute task that is fine. For an overnight reconciliation it is the whole design.

Assuming Windows on Arm runs everything. NVIDIA reports native PyTorch-CUDA, WSL and 97% of the top 15,000 Python packages working. That covers developers well. It says nothing about your accounting add-in, your label printer driver or the VPN client, and those are what stop a rollout.

Letting the agent sign in as the person. Whether the agent runs on the laptop or elsewhere, it acts with somebody’s rights. If those rights are the user’s own session, with their saved passwords, nothing tells the agent’s actions apart from theirs. That is why identity is the real lock-in again: the directory has to know about agents too.

Leaving the work on the device. An agent’s history, memory and half-finished jobs are company records. If they live only on the laptop, they leave with the laptop at the end of the lease, or with the person who leaves the company.

What to do instead

Split the fleet by role, not by budget. A handful of people who build or test models, or work offline with sensitive data, have a real case for a laptop that runs agents locally. Everyone else needs a good ordinary laptop and access to an agent that runs somewhere you manage.

Put long-running agents on something that stays on. That can be a desktop station on your own premises, a server, or a managed workspace in the sense NVIDIA’s reference design uses. The test is the same in each case: the job carries on with the lid closed, and an administrator can see it, pause it and take it away.

Make agent identity a directory decision. One account per person, with the agent acting through a broker that holds the credentials, not the agent process itself. Removing someone should remove their agents’ access in the same step.

Own the update path. A local agent stack has a model, a runtime, drivers and an operating system, each updated on its own schedule. Someone has to know which version each machine runs and decide when it changes, exactly as with the rest of the fleet.

Write down what stays with the company. Agent logs, memory and outputs belong in company storage, not in a user profile. Then the hardware can be swapped without losing the work, which is what lets a laptop outlast its lease or be replaced on schedule without drama.

Pilot both before the refresh. Two of the expensive laptops, two ordinary ones pointed at a managed agent, the same three tasks for a month. The answer will be about your work, not a benchmark.

Which of the two is right for a given company, time will tell; the hardware is weeks old. The criteria will not change: identity, updates, and what stays on your side. That is the same audit we start every engagement with at SovereOS.

A test for this afternoon

Pick one task you would like an agent to do: chasing unpaid invoices, sorting supplier mail, preparing the weekly report. Ask three questions and write the answers down.

Where does it run at 22:00, when everyone’s laptop is closed? Under whose account does it act, and how would you remove that access tomorrow? Where is the record of what it did, and does it survive the laptop being returned?

If any answer is “on the user’s laptop”, you have found what the next purchase has to decide.

Sources: NVIDIA, RTX Spark availability, 16 October 2026, NVIDIA, Secure Agent Workspace reference design.