By Cursor 3.5 shipped May 20, and it reframes a question developers have been asking for the past two years. The question used to be “how do I use AI to write code faster?” With Cloud Agents, the question is “which parts of my backlog can I just hand off?” These agents run in isolated cloud VMs, build and test the actual software, and submit merge-ready pull requests — while your laptop is closed.
What Cloud Agents Actually Are
Cloud Agents are not a smarter autocomplete. They are autonomous agents that run inside dedicated virtual machines on Cursor’s infrastructure. Each VM gets a full development environment: your repos cloned, dependencies installed, secrets configured, and network access available. The agent can run terminal commands, execute tests, browse documentation, and iterate on its own output. When it’s done, it delivers a pull request with screenshots, logs, and a screen recording showing the software actually running.
You configure the environment via a environment.json file in your .cursor directory — point it at a Dockerfile, use a saved snapshot, or let Cursor inspect your repos and generate the config automatically. The single biggest lesson from Cursor’s own post-mortems: cloud agent failures trace back to environment problems, not model problems. An agent without a working dev environment is like a developer without a computer.
Three Workflows Worth Learning
Plan Mode
Draft a plan in your local editor. When it looks right, hand it to a Cloud Agent for implementation. You move on to the next problem while the agent builds. When it finishes, you review a PR instead of a prompt.
The Race Pattern
Spawn three agents on the same task in three isolated worktrees. Each takes a different approach. You compare diffs and merge the winner. This changes the economics of exploration: three approaches costs the same amount of your attention as one. It’s particularly useful for performance-sensitive code or architecture choices where the right answer isn’t obvious until you see it running.
Automations
Cloud Agents can be triggered from outside the IDE entirely. A Slack message drops into your engineering channel; an automation investigates the root cause, checks for duplicate issues, creates a Linear ticket, and replies in the thread — no developer touched a keyboard. Supported triggers include GitHub, GitLab, Linear, Slack, PagerDuty, cron schedules, and custom webhooks. Rippling’s engineering team uses this for incident triage, weekly status reports, and on-call handoffs.
Cloud Agents vs. Local: An Honest Framework
Use cloud for well-scoped tickets with acceptance criteria and an existing test suite. Use local for architectural decisions, ambiguous problems, and anything requiring real-time back-and-forth. The agent uses your tests as its feedback loop — it runs them, sees failures, and iterates. No tests means no feedback loop means expensive noise.
Cursor reports a 30%+ PR pass-and-merge rate for Cloud Agents, and organizations using Agent Mode broadly see 39% more PRs merged overall. Those numbers assume decent test coverage and well-written tickets.
The Cost Is Real
Cloud Agents are available on the Pro plan ($20/month) and above, but they bill separately from your subscription credits and require MAX mode, which adds a 20% surcharge on every run. At scale, this adds up fast. One developer reported the same task consuming 80% of their Cursor daily credits versus 12% when run through a CLI agent. Whether the productivity gains justify the cost depends entirely on how well-defined your work is.
What This Actually Means
The engineering teams getting the most out of Cloud Agents share one trait: a groomed backlog with clear acceptance criteria and solid test coverage. If you have that, you can run your ticket queue overnight and review PRs in the morning. If you don’t, Cloud Agents will confidently produce plausible-looking code that fails in production. The technology is real. The prerequisite is discipline, not just a $20 subscription.
The full Cloud Agents announcement and official documentation are on Cursor’s site.
