AgentsCloud agent computers
Some agents don't just answer. They work in the cloud: on a computer of their own, on a browser or sandbox they borrow for each task, or as an automated workflow. Here's what that means, when it beats running an agent on your own machine, and what a business should check before turning one on.
✓ Checked 2026-10-15 against official vendor pages
Three ways an agent can run in the cloud
"Cloud agent" gets used for three very different things. Knowing which one you're buying tells you what the agent can do, what it keeps between tasks and how it's billed.
- Cloud computer. The agent gets a real computer of its own in a data center: a full desktop or virtual machine with apps, a browser and files, dedicated to your agents and kept for you (or checked out from your own managed pool). It can use desktop apps, not only websites, and it's still there tomorrow. Examples: Windows 365 for Agents, Grok Bot's cloud computer and Manus Cloud Computer.
- Per-task cloud browser or sandbox. The agent borrows a browser, VM or code sandbox for one task or conversation, and then it's reset or thrown away. Good for web errands, research and coding tasks. You don't get a computer you can come back to. Examples: the cloud browser in ChatGPT Work, Codex Cloud, Claude Code cloud sessions, Gemini Spark, Jules, Copilot Studio computer use and Researcher with Computer Use.
- Cloud-hosted automation. No screen at all. The agent is a workflow (trigger, AI step, action) that runs on the vendor's servers. Examples: n8n Cloud, Zapier and Make. See cloud-hosted automation below.
In every case you hand the agent a task, close the lid, and it keeps going. When it needs a password or your approval, it pauses and asks.
What each major lab offers
Only two of the five major labs sell a true cloud computer for agents today. The rest offer per-task cloud browsers and sandboxes.
- OpenAI: no true cloud computer for agents. Per-task: ChatGPT Work: cloud browser (GA); Codex Cloud (GA).
- Anthropic: no true cloud computer for agents. Per-task: Claude Code cloud sessions (GA); Claude computer use (API) (GA).
- Google: no true cloud computer for agents. Per-task: Gemini Spark (Rolling out); Jules (Preview).
- xAI: cloud computer: Grok Bot (GA).
- Microsoft: cloud computer: Windows 365 for Agents (GA). Per-task: Copilot Studio computer use (GA); Researcher with Computer Use (Microsoft 365 Copilot) (GA); GitHub Copilot cloud agent (GA).
Microsoft, plainly: Windows 365 for Agents is Microsoft's cloud PC product for agents (generally available, from $0.40 per Cloud PC hour in the US). Autopilot in Microsoft Copilot, announced on September 25, 2026 and previously called Scout, is the persistent agent to watch: Microsoft says it lives in your tenant with "its own identity, memory, computer and workspace" and you can @mention it in Teams and Outlook. It's in private preview with usage-based billing and no published price, and Microsoft hasn't said publicly whether its computer is a Windows 365 for Agents Cloud PC. Copilot Studio computer use and Researcher with Computer Use borrow a hosted browser, sandbox or pooled Cloud PC per task, so they're in the per-task group. Don't confuse any of this with Windows Autopilot, an older Intune service that sets up new company laptops.
Full details for every product, including AWS, Perplexity, Cursor, Devin, E2B, Browserbase and Cua, are in the cloud agent computers comparison.
Cloud or local: how to choose
| Question | Cloud agent computer | Agent on your own machine |
|---|---|---|
| Keeps working when you log off | Yes. Tasks continue after you close your laptop. | Only while your computer is on and awake. |
| Risk to your own device | Low. The agent works on a separate machine, often reset after each task. | Higher. It can reach your files and signed-in accounts unless you lock it down. |
| Where data goes | To the vendor's cloud (or your Azure or AWS account for builder platforms). | Can stay on your hardware, especially with a local model. |
| Cost model | Included in a plan, or metered per hour, per step or per vCPU-second. | Your hardware and electricity, plus model costs. |
| Admin control | Strongest on enterprise offerings, with identity, device policy, audit and domain allow lists. | Depends on how you manage the device. |
| Best for | Long or overnight tasks, many tasks at once, and teams that need governance. | Private documents, offline work and tinkering. See local AI agents. |
Business checklist
1. Security
- Who does the agent sign in as? The best answer is its own identity, not a person's password. Windows 365 for Agents gives each agent its own Entra agent identity. ChatGPT's cloud browser uses a sign-in form whose contents the model can't see.
- Is the machine reset between tasks? Pooled Cloud PCs reset when checked back in, and Researcher's sandbox is deleted after each conversation. Persistent cloud computers, such as Grok Bot's or Manus Cloud Computer, keep files and logins, which is convenient but means anything stored there is reachable by every agent on that account.
- Prompt injection. Web pages can hide instructions meant for the agent. Keep approvals on for payments, sending and deleting, limit the sites it can visit, and start on sites you trust.
2. Compliance and data residency
- Check the region. Windows 365 for Agents lets you choose an Azure region in the US, Canada, Europe and the UK, Asia, India, Australia, Africa, Brazil or the UAE. Browserbase runs in the US, the EU and the UK. Nova Act runs in US East.
- Check the contract. For example, OpenAI says Codex in the cloud isn't covered by its BAA, so it's not for patient data. Browserbase offers a HIPAA BAA on its Scale plan.
- Check training. Business and enterprise plans usually exclude your data from model training. Consumer plans often need you to opt out.
3. Cost
- Included in a seat: the cloud browser in ChatGPT Work, Claude Code cloud sessions (no separate VM charge), Grok Bot (X Subscription Premium and up), Google's Jules (daily task limits) and the GitHub Copilot cloud agent (uses your Copilot plan's AI Credits).
- Metered: Windows 365 for Agents ($0.40 per Cloud PC hour in the US), Copilot Studio computer use (5 Copilot Credits per step), AgentCore Browser ($0.0895 per vCPU-hour), Nova Act ($4.75 per agent hour) and Browserbase ($0.10 to $0.12 per browser hour after included hours).
- Measure cost per completed task, not cost per seat. A run of 20 steps in Copilot Studio uses 100 credits, or 300 with a premium model.
4. Admin controls to ask for
- An on/off switch and pilot groups (Researcher with Computer Use, and workspace policies for ChatGPT Enterprise and Codex)
- Site allow and block lists, and app allow lists for desktop computer use
- Device policy and audit through Intune, Defender and Purview for Windows 365 for Agents, or IAM and CloudWatch on AWS
- Spending limits, using Azure billing policies, AWS budgets and Copilot credit caps
Cloud-hosted automation (not a cloud computer)
Some products run your agent as a workflow on their servers: a trigger, then AI or app steps, then an action. That is cloud-hosted automation. It is useful and widely used. It is not the same as giving an agent its own desktop or browser to look at and click. If the agent never sees a screen of its own, it belongs here, not in the tables above.
When to choose which
| Choose… | When it fits |
|---|---|
| Cloud-hosted automation (n8n Cloud, Zapier, Make) | The job is a known sequence across apps you already use: "new lead → research → draft email → update CRM." You want reliability, logging and a clear cost per run. The agent does not need to open random websites or desktop apps. |
| Cloud agent computer | The job needs a real screen: clicking through a supplier portal, using a Windows app with no API, or overnight research in a browser the agent owns. See the providers above and the comparison. |
| Local agent or local model | Data must stay on your machine, you are offline, or you are tinkering. See local AI agents and run models on your own computer. |
n8n Cloud, Zapier and Make
| Product | How it works | AI / agent features | Price (checked 2026-10-15) | Self-host? |
|---|---|---|---|---|
| n8n Cloud n8n | You build a visual workflow: a trigger (new email, form, schedule), then steps that call apps, run code or ask an AI node, then an action (Slack message, CRM update). n8n Cloud hosts the instance so you do not manage servers. One run of the whole workflow counts as one execution, no matter how many steps it has. | AI Agent node, MCP client/server, and other AI nodes are included on Cloud plans. Assistant (preview) helps you build workflows and comes with monthly Assistant credits (1,600 on Starter; up to 9,600 on Pro). You can also use AI models through n8n without bringing your own API keys. | Cloud Starter $24/mo monthly or $20/mo billed annually (2.5K executions). Pro $60/mo monthly or $50/mo billed annually (10K executions). Business $960/mo monthly or $800/mo billed annually (40K executions; self-hosted license). Enterprise custom. Unlimited users and workflows on every plan. | Yes. The Community Edition is free on GitHub and includes AI agents and workflows. Paid Business and Enterprise licenses unlock SSO, Git version control, queue mode and other team features on your own servers. |
| Zapier (AI by Zapier / Agents) Zapier | You connect apps and build a Zap: a trigger, then one or more action steps. AI by Zapier is an AI step inside the Zap editor that can reason, call tools and pass results to the next step. Zapier Agents (agents.zapier.com) still exists; Zapier is migrating Agents into AI by Zapier so agent runs live in the same editor and history as other Zaps. | AI by Zapier steps support model tiers (Standard, Advanced, Premium), per-tool approvals and structured outputs. Copilot helps build Zaps from plain language. Agents that have not migrated yet still run at agents.zapier.com and are measured in activities on the Agents add-on. | Free $0 (100 tasks/mo). Professional from $19.99/mo. Team from $69/mo. Enterprise custom. AI by Zapier uses the same task pool as other Zap steps (no separate Agents subscription once migrated). Higher task tiers cost less per task. Zapier has not set a shut-off date for Agents yet. | No. Zapier runs only as a hosted cloud service. |
| Make Make | You build a visual "scenario": modules for triggers, filters, routers and actions across 3,500+ apps. Each module that processes data uses credits (most use 1 credit). Make hosts the scenarios on its cloud. There is no separate desktop or browser for the agent to look at and click. | Make AI Agents (open beta) are available on all plans with Make's AI Provider, or with your own OpenAI/Anthropic key on paid plans. Running an agent costs 1 credit per operation plus token-based credits when you use Make's AI Provider. Make AI Toolkit and other AI modules are also available. | Free $0 (1,000 credits/mo). At 10,000 credits/mo: Core $12/mo monthly (from $9/mo billed annually), Pro $21/mo monthly (from $16/mo billed annually), Teams $38/mo monthly (from $29/mo billed annually). Enterprise custom. Extra credits can be purchased in bundles. | No public self-host edition. Make is a hosted cloud product. Enterprise can add on-premises agents to reach local networks. |
Related agent profiles on this site: n8n, Zapier Agents and Lindy. Lindy is a no-code AI assistant for inbox and meetings, not a step-based workflow tool, but readers often compare it with these.
For a wider business-automation shortlist that also includes Microsoft and Google products, see how to pick a business automation agent and business automation compared.
A safe first pilot
- Pick one repetitive web task with a clear finish line, such as downloading monthly invoices from three supplier portals.
- Use the cloud agent that comes with a plan you already pay for, and keep confirmations on.
- Give it access to only the sites it needs. Sign in through the vendor's secure sign-in flow, and never paste passwords into chat.
- Run it five times while watching, then measure time saved and cost per run before you widen access.
Next: compare every provider side by side in the cloud agent computers comparison, or see how to pick a business automation agent.