---
title: "Where should your agents run: PC, station, workspace, factory?"
url: "https://predictes.com/blog/where-should-your-agents-run"
description: "October brought four places to run an AI agent: a PC, a desk station, a managed workspace and an AI factory. Choose by duration, data, control and cost per task."
---

# Where should your agents run: PC, station, workspace, factory?

October 29, 2026·5 min read·predictes

A finance team pilots an agent that reconciles supplier statements against the ledger. It runs on the controller’s new laptop, it works well, and on Friday evening the controller closes the lid. The agent was halfway through four hundred statements. On Monday nobody can say which ones were done, because the only record of the run was on a machine that was asleep for the weekend.

Nothing in that story is a model failure. The agent was put in the wrong place for its job.

## Four places, all announced in one month

That question has more answers than it did a year ago. In October NVIDIA announced, or started shipping, four quite different homes for an agent.

**The PC as a teammate.** [RTX Spark laptops](https://nvidianews.nvidia.com/news/nvidia-rtx-spark-pcs-available-today-a-new-beginning-for-windows-pcs) went on sale on 16 October: Windows 11 on Arm with a single superchip and, according to NVIDIA, up to 128 GB of unified memory. Agents run locally, with permissions handled by NVIDIA OpenShell and Microsoft Execution Containers. NVIDIA’s own phrase is that the PC turns “from a tool into an everyday teammate”.

**The station on the desk.** DGX Spark is the Linux box that stays on: it does not travel, does not sleep when a meeting ends, and is listed with RTX Spark and DGX Station as an edge device in NVIDIA’s [OpenShell reference architecture](https://nvidianews.nvidia.com/news/openshell-reference-architecture).

**The managed workspace.** In the same architecture, the [Secure Agent Workspace](https://docs.nvidia.com/enterprise-reference-architectures/secure-agent-workspace-reference-design/latest/abstract.html) describes an always-on agent working for hours on an employee’s behalf in a managed remote environment, with the user’s device reduced to a presentation layer. Its key line is that “the workspace ≠ the session”: the agent’s work outlives the laptop’s.

**The AI factory.** [DSX POD](https://nvidianews.nvidia.com/news/nvidia-introduces-dsx-pod-ai-factories-to-accelerate-europes-infrastructure-deployment) is a reference design for complete factories on Vera Rubin, built in modular containers. NVIDIA says installation on a prepared site can shrink from six months to as little as fourteen days, with deliveries from early 2027. In Poland the EuroHPC factories [PIAST](https://eurohpc-ju.europa.eu/ai-factories/poland_en) in Poznań and Gaia in Kraków are the nearest public equivalents.

These are complementary, not competing. The mistake is choosing among them by which announcement arrived first.

## How the placement decision goes wrong

**The hardware is chosen before the task.** Somebody buys the machines the demo ran on, and the agents are then fitted to them. How long the agent works, and whether it must keep working when its owner leaves, is asked afterwards, if at all.

**The session is mistaken for the job.** A chat assistant lives as long as the window is open. An agent that reconciles, monitors or prepares a month-end pack does not. If its state, its log and its credentials live in one person’s session, the job stops when the session does — as on Friday evening.

**Data goes wherever the agent happens to be.** A local agent pulls customer files onto a laptop because that is where it runs. A remote one pulls the same files into a region nobody chose. In both cases the data’s location was decided by the agent’s location, which is backwards.

**Control has no named owner.** On a laptop, the user is implicitly in charge of the agent’s permissions. In a workspace it is IT. In a factory it is whoever operates the factory. When nobody writes this down, every incident starts with an argument about whose problem it is.

**The cost is counted per device, not per task.** A laptop price and a cloud bill are both easy to read, and neither says what one finished task costs, which is the only one that should decide.

## Decide with four questions instead

1.  01How long does it work?Minutes beside a person, or hours and days on its own, including when nobody is logged in.
2.  02Where does the data have to stay?Move the agent to the data, not the data to wherever the agent was installed.
3.  03Who owns control?A named person or team for permissions, policies, logs and revoking access.
4.  04What does one finished task cost?Compute, approvals, escalations and upkeep, compared across the places that pass the first three.

The first question usually rules out half the options on its own.

**Duration decides first.** An agent that works with a person, on that person’s files, for minutes at a time, belongs close to them: a PC or a desk station is a reasonable home, and local inference keeps the loop fast. An agent that works for hours, overnight, or across many people’s requests does not belong on anything with a lid. It needs a place that stays up when its owner leaves — a station that never sleeps for a small team, a managed workspace for the many.

**Data decides second.** Put the agent where the data already is allowed to be. Sensitive records that may not leave a site argue for local or on-premises compute. Data already in a governed cloud tenant argues for a workspace in that tenant. Whatever leaves the agent’s environment is a separate question, covered in [an agent sandbox is mostly about what can leave it](https://predictes.com/blog/an-agent-sandbox-is-mostly-about-what-can-leave-it).

**Control decides third.** The Secure Agent Workspace design is useful here even if you never use its hardware, because its maturity model reads as a list of who must own what. Phase 0 is the baseline: configuration, patching, managed images, telemetry to the security team, rebuild and revoke. Phase I controls the perimeter: one user’s agent on an approved image, single sign-on, an allow-list of destinations, human review of writes. Phase II enforces around the agent itself: default-deny egress, signed policies at each tool call, brokered credentials. NVIDIA calls the phases “additive control outcomes, not deployment-status labels”. If nobody in your organisation can own Phase I, an always-on agent is not yet a good idea anywhere. What the tools themselves should allow is a design question of its own — [an agent is only as safe as the tools you gave it](https://predictes.com/blog/an-agent-is-as-safe-as-its-tools).

**Cost decides last, per task.** Only now compare money, and compare it in the unit argued in [an agent’s business case is a cost per task](https://predictes.com/blog/an-agents-business-case-is-a-cost-per-task-not-an-autonomy-rate). A factory matters to most companies indirectly: it is where the hosted models and the workspaces of the next few years will run. Futurum [points out](https://futurumgroup.com/?p=97024) that in Europe the binding constraint is access to grid power, which a faster build does not solve. Treat capacity and its price as assumptions with a date on them, not as givens.

## A test you can run today

List the agents you run or plan to run this quarter. For each one, write four words: how long it works, where its data lives, who owns its control, and whether its cost per task is known. Then ask one more question of every agent on a personal device: what happens to its work when the owner closes the lid? If the answer is “it stops” and the job should not, you have found your first placement decision.

## Our part

In [AI managed services](https://predictes.com/services/ai-managed-services) we run agents after go-live: the logs, the policies, the evaluation set and the cost per task, wherever the agent is placed. We do not resell hardware, workspaces or capacity, so the placement is your choice and we will say when a laptop is enough. More about predictes is on the [home page](https://predictes.com/).

-   AI agents
-   infrastructure
-   agent workspace
-   AI PC

## Keep reading

-   [One source, three stages: detecting an autonomous attack](https://predictes.com/blog/one-source-three-stages-detecting-an-autonomous-attack)
-   [How to build an agent factory: the loop, stage by stage](https://predictes.com/blog/how-to-build-an-agent-factory-the-loop-stage-by-stage)