Cursor Cloud Agents Now Author 56% of Its Own Merged Pull Requests

Clarifying the 56% figure

The number spreading through developer channels this month is easy to misread. Cursor's cloud agents do not merge 56% of all pull requests across the industry. They produce 56% of the pull requests merged inside Cursor's own company. In a statement posted to its official account, the company said that in December one in ten of its merged PRs came from cloud agents, and that the share is now 56% because it uses those agents to finish longer engineering tasks from start to finish1. This is an internal, self-reported measure from the company selling the product. It still tells us a lot about where agentic coding is heading, but it is not an independent count of how AI is reshaping software teams in general.

The trajectory is the more useful signal. When Cursor launched cloud agents with computer use on February 24, it said 35% of its internal merged PRs came from agents running on their own in cloud sandboxes2. A Cursor team member described the figure on Reddit at the time as roughly 30%4, and other posts from launch week repeated the 30% number7. Whichever February baseline you accept, the share has gone from about one in ten to more than half in under a year. The company ties that jump directly to infrastructure it built for unattended agent work1.

What Cursor shipped in October

The new figure arrived with a quick run of releases. On October 6, Cursor added the ability to control agents running on a developer's own computer from its iOS app1. On October 13, it launched Origin, an app marketplace with more than 20 integrations including Cloudflare and Linear, and opened it to public submissions the same day1. On October 16, cloud agents gained the ability to build interactive canvases that can be shared by link. Cursor also released an Environments API in beta and raised cloud environments to 300 repositories, with large environments said to load five times faster1.

Those releases follow September's Projects beta. Projects puts a coordinator agent on a cloud machine, where it hands work to subagents and can watch Slack channels, follow PRs, fix CI, or run on a schedule without being prompted9. Cursor says new Projects users merge 30% more PRs and heavy users merge six times as many. Coverage has been careful to call those figures self-reported and unaudited10. One internal Projects setup is reportedly on track to touch 20 to 100 PRs a day10.

Taken together, the releases explain the 56% figure. Cursor is building a system where agents keep running when the laptop closes, reach outside the repository, and wake up when something happens elsewhere. A rising agent share of merges is what you would expect from that design.

How good the merged PRs are

A high share of merges does not prove the merged code is good. Outside data is limited but somewhat reassuring. A January arXiv study of 8,106 fix-related PRs from five coding agents found 68.4% of Cursor's PRs in the sample were merged. Codex had the highest rate at 81.6% and Devin the lowest at 42.9%8. The authors also found that when Cursor and Claude Code PRs were closed without merging, the reasons leaned toward prioritization or duplicate work rather than consistent technical errors8. Cursor's sample was small, though, at 411 PRs8.

Cursor's own Insights data adds detail. In human-started cloud-agent conversations, longer exchanges were much more likely to open a PR: about a third of one-turn conversations did, compared with more than four in five conversations of 12 turns or more. The share of those PRs that merged fell slightly, from 67% to 59%3. Over the longer term, a University of Chicago study Cursor promoted in 2025 found companies merged 39% more PRs after agent mode became the default, with no significant change in revert rate5.

None of this settles whether more output means proportionally more value. Skeptics in the community asked how many agent-authored PRs needed a second pass, and argued that an agent's decisions and quiet simplifications end up in logs no one reads4. That concern about accountability is the right lens for the 56% figure.

Code review is now the bottleneck

When agents write most of the code, reviewing it becomes the hard part. Analysts made this point when cloud agents launched: one developer directing 10 to 20 parallel agents multiplies the number of PRs waiting for review2. October's releases across the industry suggest vendors agree.

Anthropic's Claude Code v2.1.304, released October 19, added organization-level Code Review settings with Quick, Standard, and Deep levels that can be set per repository. It also added an option to include every repository in a GitHub organization automatically16. GitHub made agentic autofix for code-scanning alerts generally available. Copilot explores the codebase, proposes a fix, and reruns CodeQL to confirm the alert is resolved before opening a PR for human review23. VS Code's Agent Merge uses GitHub's HydraFusion orchestrator to handle review feedback and merge conflicts28.

The risks are already visible. A Copilot code review incident lasting about two hours and 45 minutes left reviews stuck in a pending state, with delays of up to two hours, and reviews in data-residency regions did not complete at all23. When review itself depends on AI, an outage in that layer stalls the whole merge pipeline. Teams leaning on agent-authored PRs should plan for that.

Claude Code ships constantly

Anthropic's approach differs from Cursor's but is converging on it. Claude Code shipped at a fast pace this month. Claude Mods, introduced October 1, let JavaScript or TypeScript plugins draw panes, intercept tool calls, and call models18. Version 2.1.302, released October 16, turned on forked subagents by default13. Earlier, v2.1.296 added a variable that pins every workflow subagent to a single model15. The tool had reached v2.1.305 by October 2019.

Much of the work is security plumbing. One changelog summary lists repeated fixes for sandbox and permission gaps, including a plugin-directory sandbox escape that was closed in 2.1.29914. This is the hidden cost of agent autonomy: every new hook or plugin surface needs its own guardrails.

GitHub Copilot competes on reach and guardrails

GitHub is playing a different game. Copilot made local sandboxing generally available on October 7 across Copilot CLI, the Copilot app, and VS Code Agent Host sessions24, at no extra charge25. VS Code 1.142 now lets administrators require that chat sessions run in the Agent Host harness, where policies can be enforced24. Copilot CLI and the Copilot app also gained a computer-use preview for graphical environments22.

On models, Copilot made Gemini 4 Argon generally available across eight surfaces on October 1325, then retired six models on October 19 in its largest cutover so far21. GitHub's main advantage over Cursor remains distribution through VS Code and its existing enterprise footprint1.

Comparisons between the two vendors do not always agree. One October comparison said Cursor offers no computer-use feature25. But Cursor's February launch was built specifically on cloud agents with computer use that test their own work by clicking through a browser2. The disagreement seems to come from GitHub's preview controlling local desktops while Cursor's agents operate inside cloud VMs. Buyers should read "computer use" claims carefully.

The takeaway

The 56% figure is a vendor's own data, and the headline version overstates it. Even so, it is the clearest public evidence yet that an engineering organization can make agents its main source of code changes. External merge-rate studies suggest agent PRs are not mostly junk8, but they don't measure review burden or downstream quality at Cursor's scale.

The competition is no longer mainly about autocomplete or benchmark scores. It is about who controls the loop of agent execution, review, and governance. Cursor is betting on cloud orchestration and an open marketplace1. GitHub is betting on enterprise reach and enforced sandboxes24. Anthropic is betting on extensibility and fast iteration18. The open question, raised in coverage of Origin, is how actions agents take through third-party integrations get logged, rate-limited, and reversed1. Until vendors answer that, engineering leaders should treat agent-majority merges as a target to work toward carefully, not a setting to switch on.