Cursor Cloud Agents Now Run in Vercel Sandbox: Isolated microVMs, Scale-to-Zero, and a New Answer to 'Where Should Agents Execute?'

🐾 Quick facts

Free tier：There is
Cheapest paid plan：US$30/mo and up
Free quota：The free plan is US$0/month, with a quota that is a dual limit of …
Last checked：2026-09-21

Article last updated：2026-09-08

Vercel’s changelog for 2026-09-03 announces that Cursor Cloud Agents can execute in Vercel Sandbox, on isolated microVMs, with scale-to-zero. Verified 2026-09-08.

Small on its own. Put it next to Cursor’s own Self-hosted machines entry from 2026-09-02 and the direction is obvious.

1. Two new execution options in one week

Date	Publisher	Where agents can run
2026-09-02	Cursor	Your laptop, a VM, a team pool — or AWS Lambda, Coder, Cloudflare, Daytona, Modal, Namespace, Vercel, E2B
2026-09-03	Vercel	Vercel Sandbox (isolated microVM, scale-to-zero)

Both describe the same shift from two sides: “where the agent executes” went from a fixed answer to a field you fill in.

2. What those two terms actually mean

Plainly:

microVM — a harder wall than a container. If the agent wrecks the environment, installs junk or deletes the wrong thing, the only casualty is one disposable machine. Your production environment is not in reach.
scale-to-zero — with no task running, the machine count is zero, so nothing idles on your bill. Capacity appears when work does.

Why they are quoted together becomes clear once you look at the workload shape: agent work is bursty. Nothing all day, then ten tasks at once. The worst fit for that shape is a machine idling around the clock.

3. Choosing between this and self-hosting

The practical call. Three placements, each with its case:

Placement	Good for	Bad for
Vendor cloud (default)	Just using it, no infrastructure to run	Rules that forbid data leaving your environment
Your own machine	Complex environments (packages, credentials, intranet); recycling an idle laptop	Nobody to operate it; agents touching real files
Isolated sandbox like Vercel’s	An environment that can safely be destroyed; bursty work you do not want to host	Tasks needing intranet or local file access

The point most easily missed: an isolated sandbox is not about speed, it is about being allowed to break. The expensive part of agent work is usually not compute — it is your time cleaning up what it broke.

4. What is still unknown

Stated plainly:

The Vercel entry publishes no pricing detail for this combination — how Sandbox usage is metered, and how it splits from your Cursor subscription, is not written.
Nor whether any free allowance exists for trying it.
Neither side gives performance or cold-start numbers.

Check each vendor’s usage and billing pages before committing. Do not read the announcement’s tone as “free” or “cheap”.

5. Why it matters

If you already run Cursor cloud agents, this gives you another placement, not a new capability. The trend is the part worth keeping: AI coding tools are unbundling the execution environment into a pluggable layer.

That is good for users, because it adds a question you can ask during evaluation: can this tool’s agents only run on the vendor’s own infrastructure? A tool that answers no carries less lock-in risk.

The other half of Cursor’s week: self-hosted machines and computer use. Using one Gateway key across models: AI Gateway setup.

Sources read directly on 2026-09-08: the “Cursor Cloud Agents in Sandbox” entry dated 2026-09-03 in the official Vercel changelog, and the 2026-09-02 entry in the Cursor changelog. Isolated microVMs and scale-to-zero are Vercel’s wording. Neither announcement publishes pricing, free allowances or performance figures for this combination, and we do not estimate them. The comparison table in section 3 is our framework, not vendor guidance.