Cursor says cloud agents write 56% of its merged PRs

• Cursor says cloud agents now author 56% of its merged PRs, up from one in ten in December, in Cursor's 56% post.
• The implementation detail is boring and important: Cursor's blog link points to Ubuntu cloud VMs, Dockerfile setup, secrets controls, a repo-specific CLI, and self-healing diagnostics.
• Cloud Doctor turns environment drift into PRs: ericzakariasson's Cloud Doctor post shows the hourly doctor diagnosing unhealthy agents, proposing fixes, and feeding approved patches back into the repo.
• Cursor also made the review loop mobile, with iPad support and a full PR review surface for comments, checks, approvals, and merges in Cursor's iPad launch post.
• The cloud-agent workflow is already normal enough that kylebrussell says he fires Cursor agents from his commute, while jaredpalmer says he still has not set up local development at Cognition.

Cursor's official environment writeup has the engineering meat: Mac-local dev had to be made Ubuntu-safe, secrets were wrapped with egress limits and redaction, and a tiny-sounding internal CLI named anydev removed build-command footguns. The Cloud Agents docs say each agent gets an isolated VM with a full dev environment, screenshots, videos, logs, and optional GitHub PR embedding. The iPad changelog turns the same loop into a mobile surface: pinned agent chats, split-screen review, full-file diffs, screenshot markup, and merge from anywhere.

Cursor's metric is PR share, not lines of code or suggestion acceptance rate. Its post says cloud agents moved from roughly 10% of merged PRs in December to 56% today.

The stronger part is the denominator: merged PRs. The weaker part is that PR count hides size, complexity, review effort, and whether agents are taking the hard work or the well-shaped backlog.

Cursor's environment writeup frames the dev environment as a product whose users are agents. Its setup docs define the target as an isolated Ubuntu machine with the same repos, tools, dependencies, secrets, startup commands, and network access a developer would have.

The practical pieces:

• Most Cursor developers work locally on Macs, while cloud VMs run Linux.
• Cursor agnosticized dev utilities and setup scripts for Ubuntu.
• A Cursor-defined Dockerfile became the starting image for cloud agents.
• Cloud agents can run in parallel without a local machine staying online, according to the Cloud Agents docs.
• The capabilities docs say agents can use a desktop and browser, click through UI flows, and attach screenshots, videos, and logs to PRs.

Cursor also lists four security controls around secrets and source access:

• Network egress restrictions.
• Scoped and proxied git remote access.
• Secret scanning in commits and commit messages.
• Secret redaction in tool results, including cases where the agent tries to read the value.

Cursor says skills helped document build workflows, but did not remove the footguns. The bigger move was anydev, an internal CLI that turns scattered developer commands into an agent-friendly interface in the environment writeup.

anydev does four jobs:

• Starts all services.
• Routes common utility scripts.
• Provides multiple --help menus for subcommands.
• Supervises long-running build commands and restarts them, so the model does not have to babysit processes.

That supervisor is the sharp bit. Cursor moved responsibility from prompt memory into deterministic tooling.

Cursor built Cloud MCP so agents could inspect their own cloud environments with dynamically discoverable tools, according to the environment writeup. The same writeup says agents use it to diagnose setup failures, egress policy problems, changed secrets, and related environment drift.

Cloud Doctor sits on top of that loop:

• Periodically checks for failures.
• Separates transient errors from salient ones.
• Performs root-cause analysis.
• Opens PRs for high-confidence fixes.
• Inspects traces to find wrong skills, misleading commands, and slow workflows.
• Changes skills, commands, or the environment so later agents take a shorter path.

The failure modes are not hypothetical. Cursor's forum had same-week reports of runtime secrets missing from Cloud Agent VMs and intermittent GitLab failures involving invalid tokens, missing branches, and rate limits.

Cursor shipped iPad support the day before the environment post. The changelog says iPad is available on all paid plans, with iPhone and iPad both getting an inbox and full PR review.

The review surface now covers:

• Comments.
• Checks.
• Approvals.
• Reviewer changes.
• Comment resolution.
• Merging.

The iPad-specific layout keeps sidebar chats pinned, shows several agents at once, and puts review beside chat in split screen. Cursor also added screenshot comments, Apple Pencil markup, Bitbucket and Azure DevOps support, multi-PR sessions, and team switching.

The mobile launch makes more sense next to the hands-on posts. kylebrussell said he settled into firing Cursor cloud agents from the mobile app on his commute, and jaredpalmer said he still has not set up local development since joining Cognition because Slack and webapp work with Devin has been better so far, even for frontend work jaredpalmer.

Every's builders described a similar parallel-agent style in danshipper's Every thread. Austin Tedesco runs Codex on one task while Claude uses Descript's MCP in another loop for video work, and Yash Poojary uses Cursor cloud agents to test multiple models side by side before switching.

Eric Zakariasson put the vibe bluntly after using Cloud Doctor: the future of coding is in the cloud ericzakariasson's Cloud Doctor post.