The hardware can; the question is whether your setup can. A GB10 system such as the ASUS Ascent GX10 (from S$8,100) or NVIDIA DGX Spark (S$8,800) is a desktop built for sustained load, draws its power from a single 240 W adapter and keeps every prompt on your premises, which makes it a natural home for AI agents that run around the clock. What makes an agent safe to leave alone is not the box but four things around it: a sandbox that limits what the agent can touch, a scheduler that restarts it when it fails, logs you actually read, and a rule for which actions need a human to approve. This post covers each one. Prices at SourceIT, listed 28 Oct 2026.
What do "always-on" agents actually do?
An AI agent is a model that takes actions: it reads files, calls tools, browses, writes reports and hands work to other agents. Running one around the clock makes sense for jobs that arrive continuously or take hours: triaging an inbox overnight, watching a folder for new invoices, monitoring logs, summarising tickets, or running long research tasks. Our post Your Next AI Employee Might Be a Computer Sitting on Your Desk gives more examples.
A local system suits this better than a cloud API for two reasons. Cost is fixed: an agent that calls a model thousands of times a day pays per token in the cloud and nothing extra on your own hardware. And data stays in the building, which matters when the agent reads client files or personal data.
What does the GB10 software stack offer for agents?
| Component | Role | Notes |
|---|---|---|
| OpenClaw | Open-source agent framework: files, browsing, tools, long jobs | Validated by ASUS on the Ascent GX10 |
| NVIDIA NemoClaw with OpenShell | Wraps OpenClaw in policy-based sandboxes | Controls what an agent can read, write and reach on the network |
| Hermes Agent | Self-improving agent with a browser admin panel | Also validated by ASUS |
| NVIDIA PAIR | Routes inference from one NVIDIA machine to other PCs on the network | Lets laptops use the GB10 as their model server |
| Inference server | vLLM, NVIDIA NIM or Ollama | Provides the OpenAI-compatible endpoint agents call |
Based on ASUS's Ascent GX10 materials and NVIDIA's DGX Spark software documentation. Our NemoClaw and OpenClaw setup guide walks through installation.
How do you stop an unattended agent doing damage?
Start from the assumption that the agent will eventually misread an instruction. The aim is to make that mistake small, visible and reversible.
- Least privilege. Give each agent its own account with access to only the folders, mailboxes and systems its job needs. Read-only wherever possible.
- Sandbox network access. Use OpenShell policies or firewall rules to restrict which sites and internal systems the agent can reach. An invoice agent has no reason to browse the web.
- Human approval for irreversible actions. Sending external email, paying, deleting, publishing or changing permissions should wait for a person. Queue them for approval rather than blocking the agent entirely.
- Spending and rate limits. If the agent can call paid services, cap them.
- Separate test and live. Change prompts and tools on a test agent first.
The governance questions are the same ones raised in Would You Trust an AI Agent With Your Company's Credit Card?
An agent you cannot audit is an agent you cannot leave alone.
How do you keep it running reliably?
- Run agents as services. On DGX OS, which is Ubuntu-based, run the model server and each agent under systemd or as containers with a restart policy, so a crash at 3 a.m. restarts itself.
- Schedule batch jobs. Use cron or systemd timers for jobs that run at set times, and stagger them so heavy jobs do not collide.
- Health checks. Have a simple script call the model endpoint every few minutes and alert someone if it stops answering.
- Plan updates. DGX OS and NVIDIA's stack update regularly and often bring performance gains. Apply them in a maintenance window, not automatically mid-job. See DGX OS updates for the GX10.
- Protect the power. A UPS lets the box ride through a blip and shut down cleanly in a longer outage. Our post Would You Trust a S$200 UPS to Protect a S$10,000 Workstation? covers sizing.
- Give it air. Place the unit with clear airflow, not inside a closed cabinet.
What should you log, and who reads it?
For each agent, log the task it received, the tools it called, the files it touched, what it produced and any action it queued for approval. Send logs off the box to wherever IT already keeps them, keep them for an agreed period and set a weekly review. The first weeks of logs are when you are most likely to catch an agent doing something sensible but unexpected, which is exactly when you want to find out.
Will one GB10 run several agents at once?
Usually yes, because most agents spend their time waiting for tools rather than generating text. Several agents can share one model server. The limit is the same as for human users: all requests share 273 GB/s of memory bandwidth, so many agents generating long answers at the same moment will slow each other down. Use a smaller, faster model for routine agent steps and reserve a large model for the steps that need it. For the concurrency arithmetic, see Can One GB10 Support 20 Employees Using AI at the Same Time?
GB10 or an RTX Spark laptop for agents?
For agents that must run all the time, use a desk unit. An RTX Spark laptop such as the MSI Prestige N16 Flip AI+ runs agents well while it is open and plugged in, and NVIDIA's launch software includes guided agent installers. But a laptop gets closed, carried and put to sleep. The pattern that works is a GB10 as the always-on agent host, with laptops calling it through NVIDIA PAIR. We compare the two in Would You Rather Have a Windows AI Laptop or a Dedicated GB10 Supercomputer?
Buying in Singapore
SourceIT supplies the ASUS Ascent GX10 (1 TB S$8,100, 2 TB S$8,300, 4 TB S$10,635), the NVIDIA DGX Spark (S$8,800) and other GB10 systems with local warranty and a GST invoice, and can deliver a unit commissioned with your inference server and agent framework. Browse the Ascent GX10 collection, the ASUS Agent Computer page or the NVIDIA DGX Spark collection. For a quotation, use Request for Quotation on the product page, call (65) 6978 3502, email Sales@sourceit.com.sg or use our contact page.
Frequently asked questions
Can a GB10 system run AI agents 24/7? Yes. It is a desktop built for sustained load on a 240 W adapter. Run the model server and agents as services with automatic restart, health checks, logging and a UPS.
How do I stop an AI agent doing something harmful? Give it least-privilege access, restrict its network reach with sandbox policies such as NVIDIA OpenShell, require human approval for irreversible actions, and review its logs regularly.
What agent software runs on the ASUS Ascent GX10? ASUS validates OpenClaw and Hermes Agent and ships an optimised NVIDIA NemoClaw setup, which wraps OpenClaw in OpenShell security sandboxes.
Can several agents share one GB10? Yes. Agents spend much of their time waiting for tools, so several can share one model server, but heavy simultaneous generation slows all of them.
Is a laptop good enough for always-on agents? For agents that run while you work, yes. For agents that must run overnight and unattended, a desk unit such as a GB10 is more reliable, with laptops calling it over the network.
SourceIT Approved
Researched and verified by the SourceIT team for Singapore buyers.
