Cloud browser, sandbox or cloud computer? How to choose infrastructure for an AI agent
#Cloud browser vs sandbox vs cloud computer: the short answer
There are three shapes of infrastructure for an agent that uses a computer:
- A cloud browser fits web automation: many sessions, each one a browser, with a saved profile to stay signed in.
- A code sandbox fits running code that should start clean, or pick up exactly where a paused run stopped.
- A persistent cloud computer fits an agent that needs logins, files and desktop apps together, run after run.
Two questions decide it: what has to survive between runs, and whether the agent ever leaves the browser. If you haven't thought about the first question yet, start with Sandbox or computer? What an AI agent needs to keep between runs.
#What each option keeps and can do
| Need | Cloud browser (Browserbase, AgentCore Browser) | Code sandbox (E2B, E2B Desktop) | Windows Cloud PC (Windows 365 for Agents) | Persistent Linux computer (Burrowbox) |
|---|---|---|---|---|
| What survives between runs | The browser profile: cookies, local storage | Filesystem and memory, if you pause it | Not covered here | Disk: files, apps, vault, cookies; windows and tabs reopen. Not memory |
| Desktop apps | No | Yes, with E2B Desktop (Linux, Xfce) | Yes, Windows | Yes, Linux |
| Watch or take over live | Yes | Screen streaming (E2B Desktop) | Human-in-the-loop controls | Yes, embeddable, watch-only or interactive |
| Egress location | Not covered here | Not covered here | Not covered here | Pick a country per machine |
| Isolation unit | A session (a microVM per session on AgentCore) | A sandbox | A Cloud PC checked out per task | A machine per customer, with its own token |
"Not covered here" means we didn't check that cell for this post, and the product may well have it. The sources for each column are in the sections below.
#Where each option is strongest
#Browserbase
Browserbase runs browsers for agents. Its Contexts keep cookies, localStorage, IndexedDB, service workers and site preferences across sessions when you start a session with persist: true. The HTTP cache isn't kept (Browserbase, Contexts). It also has live view and session replays (Browserbase, What is Browserbase). If the job is the web and only the web, this is the simplest model: open a session, load the context, work, close.
#Amazon Bedrock AgentCore Browser
AgentCore Browser starts a browser session with a timeout of 15 minutes by default, adjustable up to 8 hours, and up to 500 sessions can run at once. Each session runs in a dedicated microVM that's terminated when the session ends. Sessions can be watched in a live view and recorded to S3 (AWS, Browser fundamentals). Browser profiles store cookies and local storage, so a new session can start signed in. The profile is updated only when you call the save operation (AWS, Browser profiles). It fits teams already on AWS who want clean, short-lived browsers with the login carried over. We go deeper in AgentCore Browser gives agents a cloud browser by the session.
#E2B
E2B is a sandbox for running code. Pausing a sandbox saves its filesystem and its memory, including running processes and loaded variables, and a paused sandbox is kept with no automatic deletion (E2B, Sandbox persistence). A disk snapshot can't bring back a half-finished computation; this can. For computer use there's E2B Desktop, an open-source (Apache-2.0) Linux desktop sandbox with Xfce and screen streaming (E2B Desktop on GitHub). E2B is a good fit for running code an agent wrote, cleanly and in isolation.
#Windows 365 for Agents
Microsoft's option is a class of Windows Cloud PCs for agents. An agent checks out a Cloud PC for a task and checks it back in afterwards. The machines are Intune-managed and governed by Microsoft Entra ID, with session logs and human-in-the-loop controls (Microsoft Learn, What is Windows 365 for Agents?). If the agent has to use Windows apps inside a Microsoft-managed company, this is the one built for it. We compare its check-out model with a Burrowbox pool in Microsoft checks Cloud PCs out to agents.
#When Burrowbox fits
Burrowbox gives an agent a persistent Linux computer with a desktop, a browser, a credential vault and its own MCP endpoint. It fits when one agent (or one customer's agent) has to come back to the same place:
- State across runs. Stopping a machine snapshots its whole filesystem: files, installed apps, the vault, and browser cookies including session cookies. Open windows and tabs reopen on the next start. Memory isn't kept, so programs are relaunched rather than frozen (Machines).
- Logins the agent can't read. Credentials and TOTP secrets go in the machine's vault, and the agent calls
browser_loginwithout ever seeing them (Vault). - Desktop apps. Install with
apps_install(apt, deb, AppImage, Flatpak) and drive them with desktop tools such asget_app_state,clickandtype_text(MCP). - People watching. A signed live-view link goes in an
<iframe>in your product, either watch-only or interactive, and can be cut down to one element of the page withselector(Embed the live view). - Location. A
proxyVPN location makes the browser and apps go online from a chosen country, with a stable exit IP per machine, and agents can check it withnetwork_info(VPN locations). - Clean machines when you want them. A warm pool hands out brand-new machines, and a claimed machine is never reused. Earlier posts use pools to build a per-task computer like Manus and parallel browsers like Project Mariner.
Creating a machine for one customer looks like this:
curl -X POST https://burrowbox.dev/api/machines \
-H "Authorization: Bearer $BURROWBOX_KEY" -H "Content-Type: application/json" \
-d '{"name": "acme", "size": "small", "browser": "full", "externalId": "cus_8f2a",
"ttlMinutes": 30, "vpn": {"type": "proxy", "country": "de"}}'
The response includes mcpUrl and an mcpToken that only reaches this machine. Give that token to the customer's agent. After 30 minutes the machine stops and keeps its state. A small machine costs $0.11 an hour while running and $0.001 an hour stopped, and the VPN location is billed on top (Billing).
#When it doesn't
- Windows apps. Burrowbox machines are Linux. If the agent needs Excel for Windows or a Windows-only line-of-business app, use a Windows Cloud PC.
- GPU work. Machines have no GPU. The
fullbrowser renders WebGL on the CPU, which is fine for maps and charts but slow for heavy 3D (Machines, Browser mode). - Run-once code. If the agent writes a script, runs it and throws the result away, a code sandbox is simpler. E2B also keeps memory when it pauses, which helps when a long computation has to stop and resume.
- Browser-only work at volume. If nothing outside the browser matters, a browser service with saved profiles is less to manage than a computer per task.
Burrowbox isn't involved with Browserbase, E2B, Amazon Web Services or Microsoft.
#Key takeaways
- Pick by what must survive between runs: a browser profile, a paused sandbox with its memory, or a whole disk.
- If the agent ever leaves the browser for files or desktop apps, a browser service alone won't cover it.
- Windows apps mean a Windows Cloud PC. Logins, files and Linux apps for each customer mean a persistent machine.
#FAQ
#Do I need a cloud browser or a sandbox for my AI agent?
If the agent only browses, a cloud browser is enough. If it runs code it writes, use a code sandbox. If it does both and also needs files or desktop apps that last between runs, use a persistent computer.
#Can a cloud browser keep an agent signed in?
Yes, through a saved profile. Browserbase Contexts and AgentCore browser profiles both carry cookies and local storage into later sessions. On AgentCore you have to save the profile explicitly before the session ends.
#Can an AI agent use desktop apps in a sandbox?
Yes, with a desktop sandbox such as E2B Desktop, or with a cloud computer. On Burrowbox the agent installs the app with apps_install and drives it through the machine's MCP tools.
#Is a persistent computer more expensive than a sandbox?
It depends on how long it runs. On Burrowbox you pay per minute while a machine runs and a small hourly rate for its disk while it's stopped, so turning machines off when idle matters most. We don't compare prices across vendors here.
To try it, follow the quickstart or create an account.
#Sources
- Browserbase, Contexts: https://docs.browserbase.com/features/contexts
- Browserbase, What is Browserbase: https://docs.browserbase.com/introduction/what-is-browserbase
- AWS, AgentCore Browser fundamentals: https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/browser-resource-session-management.html
- AWS, AgentCore browser profiles: https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/browser-profiles.html
- E2B, Sandbox persistence: https://e2b.dev/docs/sandbox/persistence
- E2B Desktop: https://github.com/e2b-dev/desktop
- Microsoft Learn, What is Windows 365 for Agents?: https://learn.microsoft.com/en-us/windows-365/agents/introduction-windows-365-for-agents
- Burrowbox docs, Machines: https://burrowbox.dev/docs/machines