Cursor launches cloud development environments with rollback and scoped secrets

• Cursor shipped cloud development environments for its cloud agents, so agents can run inside preconfigured VMs with cloned repos, installed dependencies, and toolchain credentials, according to cursor_ai's launch post.
• The new setup adds reusable multi-repo environments, which cursor_ai's thread framed as one environment shared across agent sessions instead of repo-by-repo setup.
• Cursor also added per-environment version history with rollback, audit logs, scoped egress, and isolated secrets, per cursor_ai's security and rollback post.
• The company says customers including Decagon, Amplitude, BILT, and Snyk already use these environments for end-to-end agent tasks, according to cursor_ai's customer examples.
• Early reaction was positive on setup quality, but olvrgln's hands-on note also reported hanging agents, while olvrgln's follow-up called out Git-centric UI assumptions that break with Jujutsu.

You can jump straight to Cursor's blog post, skim the launch thread for the laptop-style setup model, and watch the rollback and secrets demo for the admin controls. One useful detail from sjwhitmore's internal context is that Cursor had already been using similar VM setups internally, then spent time sanding down the long tail of external customer environments.

Cursor's pitch is simple: give the agent the same kind of prepared machine a human engineer would expect. That means repos are already cloned, dependencies are installed, and credentials are present before the run starts, as cursor_ai's launch post describes.

In sjwhitmore's post, the team says Cursor had already been running its own code in VMs internally and trusted the resulting work enough to merge it, then productized that setup so customers do not have to assemble the infra themselves.

The concrete workflow change is multi-repo support. Cursor says one environment can now hold all the repos an agent needs and be reused across sessions, which is a meaningful fix for tasks that span services, SDKs, and infra repos instead of living in one checkout, per cursor_ai's multi-repo thread.

That lines up with the user reaction in jediahkatz's post and sjwhitmore's retweet of dwetterau, both of which focus on not having to keep telling the agent where the other repos live.

Cursor bundled four environment controls into the launch:

• version history per environment
• rollback to an earlier state
• an audit log of changes
• per-environment egress and secrets scoping

All four come from cursor_ai's thread, which also says secrets from one environment are not accessible from another.

This is the part that makes the feature feel less like disposable demo infra and more like shared agent workspace plumbing. The launch post itself is linked from Cursor's official blog.

Cursor already had visible external usage before this launch. In ryolu_'s post, Cursor Cloud agents were used to ship several ryOS features, including TV channel surfing, a Maps app, and an iPod UI with Apple Music sync.

The first hands-on notes were not all smooth. olvrgln's report called the new environment setup "very good" but said cloud agents seemed to hang for an unreasonable amount of time, and olvrgln's follow-up said parts of Cursor's UI still assume Git rather than Jujutsu.

A separate complaint in olvrgln's comparison post also argued that cloud agents need tighter access to diagnostics like linters, tests, LSP output, and PR CI to fix obvious issues inside the environment itself.