Cursor Launches Cloud Agents That Code Autonomously for Days

Cursor, the AI-powered code editor that has become a favorite among developers, has launched a major update: cloud agents that can autonomously write code for days at a time, producing merge-ready pull requests complete with video recordings, screenshots, and execution logs.

Key Highlights

Cloud agents run in isolated VMs with full development environments, accessible from web, mobile, Slack, and GitHub
Long-running agents can operate autonomously for 25 to 52+ hours, generating over 151,000 lines of code
Over 30% of Cursor's own merged PRs are now created by these autonomous agents
A new plugin marketplace connects agents to tools like Figma, AWS, Linear, and Stripe

How It Works

Each cloud agent spins up in its own sandboxed virtual machine with a complete development environment. The system uses a hierarchical architecture: planners explore the codebase and create tasks, workers execute those tasks, and judge agents determine whether the output meets quality standards.

Rather than flat coordination between agents, Cursor found that this hierarchical approach — with planners spawning sub-planners for parallel work — scales far more effectively. The team also discovered that simplicity wins: removing unnecessary coordination layers actually improved performance.

Impressive Demonstrations

Cursor showcased several large-scale autonomous runs:

Building a web browser from scratch in roughly one week — over 1 million lines of code across 1,000 files
Migrating a codebase from Solid to React — 266,000 lines added and 193,000 removed over three weeks
Optimizing video rendering by rewriting a pipeline in Rust, achieving a 25x speed improvement
Building a chat platform integrated with an existing open-source tool in 36 hours
Implementing a mobile app from an existing web app in 30 hours

The AI Coding War Heats Up

The launch intensifies competition in the rapidly growing AI coding tools market. Cursor now competes head-to-head with OpenAI's Codex, GitHub Copilot, and Anthropic's Claude Code — all of which have introduced autonomous agent capabilities in recent months.

Cursor's in-house model, Composer 1.5, trained with 20x scaled reinforcement learning, now outperforms Claude Sonnet 4.5 on Terminal-Bench 2.0. The team also noted that GPT-5.2 excels at long-running autonomous tasks due to its ability to follow instructions precisely and avoid drift over extended runs.

New Plugin Ecosystem

Alongside cloud agents, Cursor introduced a plugin marketplace that bundles skills, subagents, MCP servers, hooks, and rules into installable packages. This allows agents to interact with external services — pulling analytics from Amplitude, deploying to AWS, or syncing designs from Figma — all within the same autonomous workflow.

Availability and Pricing

Cloud agents are available now for Ultra, Teams, and Enterprise plan users. Cursor has introduced two usage pools: an Auto+Composer pool with expanded limits and an API pool with model-based pricing. The new Composer 1.5 model receives 3x the usage limit compared to Composer 1.

Agent sandboxing, available cross-platform on macOS, Linux, and Windows, reduces interruptions by 40% compared to unsandboxed agents.

What This Means

The shift from AI-assisted coding to AI-autonomous coding is accelerating. When a tool's own engineering team uses AI agents to produce nearly a third of their merged code, it signals that autonomous software development is no longer experimental — it is becoming the default workflow for high-velocity engineering teams.

Source: Cursor Blog