Cursor AI Editor Pivots to Agents as Claude Code Review Hits GA

From smarter autocomplete to managing agents

Cursor started out as a code editor that predicted your next line. Ten months into 2026 it works more like a control panel for AI agents, and the review layer that checks those agents' output is now where the competition is sharpest. In the same weeks, Anthropic made Claude Code faster and launched its own managed review service. The result is a market where writing code with AI is no longer the hard part. Checking it is.

The pace of Cursor's releases explains much of this. One release tracker lists 35 Cursor releases, from version 1.6 in September 2025 to an "Origin App Marketplace" update on October 13, 2026, which added more than 20 integrations from partners including Cloudflare.1 Download mirrors had version 3.25.6 for Windows and macOS by October 16.69 Some outlets still describe the product mainly as a VS Code fork with smart completion, chat and a $20-per-month Pro tier.2 That description is accurate but out of date.

What Cursor has actually shipped

The turning point came on April 2, 2026, with Cursor 3. It added an Agents Window where developers run several agents at once across local folders, cloud machines and remote SSH hosts. The traditional editor is still there for people who want to type code themselves.37 Before that, Cursor 2.0 in October 2025 had introduced parallel agents using git worktrees.8 Version 2.4 in January 2026 added subagents, each with its own context window, plus Skills: task-specific workflows the agent loads only when it needs them.47

The second half of 2026 pushed further from the editor itself. According to the tracker:

• Cloud Agents (August): agents can pick up tasks from pull requests or Slack messages and work through them on their own.1
• Origin (August): an early beta that lets Cursor host code repositories directly.1
• Self-hosted machines (September 2): agents can run entirely inside a customer's own infrastructure.1
• Projects (September 10): long-running work that continues after the laptop is closed.1
• Rollouts and Security Review (September 23): two bots, one that watches deployments for regressions and one for security review.1
• Remote control (October 6): developers can watch and steer local agents from an iPhone.1

A comparison piece describes Projects as built around a "coordinator agent." It plans the work, hands pieces to as many subagents as the job needs, runs them on cloud machines, and keeps a shared file context in sync across devices.10

Cursor's own models matter here too. A tutorial cites a 79.8% score on SWE-Bench Multilingual for Composer 2.5, released in May. That figure is reported secondhand, not by Cursor.4 There has also been some controversy. Composer 2, released in March, was built on Kimi 2.5, an open-weights model from the Chinese startup Moonshot AI, and that only came out when users spotted it. A Cursor vice president then said about three-quarters of the final model's training compute was Cursor's own.8 Adoption claims are big but thinly sourced: one guide repeats a third-party report of 10 million active developers in January, up from about 4 million the previous June.4

Claude Code: speed and parallel sessions

Anthropic's terminal-first agent has moved in the same direction. In mid-October, the Claude Code desktop app gained nested sessions. Parallel tasks now spawn child sessions that report back to a main thread, and developers can save environments that combine folders, SSH hosts and cloud targets. The update requires a Pro, Max, Team or Enterprise plan.30 This lines up closely with Cursor's coordinator-and-subagent design, and it is hard to see that as coincidence. Both companies have decided developers will supervise groups of agents rather than chat with one.

Speed has been the other focus. Anthropic said on October 17 that p90 first-token latency for local Claude Code sessions fell from 4.07 seconds to 1.33 seconds between August and October. Cloud sessions dropped from 16.99 seconds to 6 seconds. Anthropic credited compressing conversation data and starting cloud sessions before users finish typing.25

Release volume is very high. Version 2.1.302 appeared around October 17, with client-side model selection for auto mode, faster startup when many plugins are installed, and a long list of reliability fixes.21 Earlier releases:

• made Claude Sonnet 5.5 the default Sonnet model, at $2/$10 per million tokens27
• added Claude Haiku 5.5 with a 1M-token context window24
• let Claude Code read AGENTS.md when no CLAUDE.md file is present28

That AGENTS.md change is small but telling. It makes it easier for teams that use more than one agent to share a single instructions file. One fix in version 2.1.300 even covers launching Claude Code inside Cursor's and Devin Desktop's built-in terminals.21 Many developers clearly run both tools side by side.

The record is not spotless. Users filed 35 reports of unwarranted refusals in April 2026, the most of any month, alongside complaints about Claude Opus 4.7.29 The Claude Code CLI's source code also leaked in March.29

Review is the new bottleneck

Code review is where the commercial stakes are rising fastest. One October analysis argues that reviewing a diff is a narrower, cheaper problem than generating code, which makes it the obvious next battleground now that the major models write code about equally well.17 The same analysis reports that Claude Code Review became generally available for Team and Enterprise customers on October 15. It is a managed service that uses several agents with a dedicated false-positive filter, offers quick, standard and deep modes, and never approves or blocks a pull request.17

The approaches differ in important ways:

• GitHub Copilot made its stronger "Balanced" review mode the default on September 28. Since September 1 it has also offered an opt-in preview that lets it submit approving reviews.17
• Cursor Bugbot stops short of approving. Its Autofix feature can push a fix through a Cloud Agent.17
• Anthropic has deliberately kept its reviewer advisory only.17

In practice, this is a split over how much authority an AI reviewer should have. Teams that turn on auto-approval casually take on real governance risk, as one analysis warns.17

Where the reporting on Bugbot disagrees

Cursor Bugbot is the clearest example of how unsettled this category still is, because accounts of its price, reach and quality vary widely.

Pricing. Several guides describe a $40-per-user add-on on top of a Cursor subscription.111619 Others say it moved to usage-based billing in mid-2026, at roughly $1 to $1.50 per review.1417 One analysis says that per-run figure is a historical estimate from launch, not a current fixed rate.17 The likeliest explanation is that the seat-price figures are simply out of date.

Supported Git hosts. Earlier comparisons list Bugbot as GitHub-only.1213 Another lists GitHub and GitLab.14 The most recent analysis says it covers GitHub, GitLab, Bitbucket and Azure DevOps.17 One guide describes it as an inline scanner inside the IDE, which conflicts with the pull-request reviewer described almost everywhere else.18

Quality. This is where the reports diverge most. AIMultiple's evaluators found Bugbot's analysis less detailed and less accurate than its rivals'.15 A CodePulse benchmark measured 84.2% precision for Bugbot, well ahead of CodeRabbit at 45.3% and Copilot at 35.3%. CodePulse sells a competing reviewer, so that result deserves some caution.19 An October hands-on comparison ranked Bugbot first for high-severity bugs, with OpenAI Codex close behind.17 Greptile's July 2025 test gave Bugbot 58%, Copilot 54% and CodeRabbit 44%.13

Read together, these results suggest Bugbot posts fewer comments, and those it does post are more often worth acting on. It does not catch everything. Tools that post many comments score well on raw catch rate but bury reviewers in nitpicks.19

CodeRabbit is the consensus default for teams that do not work in Cursor. Reviewers cite its support for all four major Git hosts and its flat price of $24 per developer per month on annual billing.141620 One benchmark notes that CodeRabbit's pre-merge checks can block a merge, while Bugbot and Copilot only give advice.15

What it means

Cursor and Anthropic are converging on the same model: an orchestrator that sends work to parallel agents, with a review system to check the results. Their go-to-market strategies differ. Cursor wants to own every layer: the editor, code hosting, deployment monitoring and review. That pitch is strongest for teams willing to standardize on one tool, though one cost estimate puts Bugbot plus Cursor Business at more than $1,600 a month for 20 engineers under the older seat pricing.11 Anthropic sells Claude Code as a layer that runs inside other people's editors and increasingly inside managed cloud services.2130

My reading is that the editor matters less than it did a year ago, and that review quality, pricing transparency and governance will decide who wins with enterprise buyers. On that front, nobody, Cursor included, has yet given buyers clear, stable pricing. That is the gap to watch going into 2027.