Cursor 2.0、新コーディングモデルComposerとエージェント中心UIを発表 New Coding Model and Agent Interface

Cursor has shipped version 2.0 of its AI coding IDE, headlined by Composer, an in-house coding model, and a redesigned interface that puts agents rather than files at the center of the workflow. The release reflects a broader shift in AI-assisted development from inline completion toward longer-running autonomous work.

Composer is described as a frontier-grade coding model trained by Cursor itself, said to run roughly four times faster than systems with comparable intelligence. Latency matters disproportionately for agent loops, where each step often involves reading code, editing files, and running tools, so a faster model can turn what would otherwise be a multi-minute wait into something closer to an interactive conversation. Independent benchmarks and details on the training setup will be needed to evaluate how Composer compares with frontier models from Anthropic, OpenAI, or Google.

The most visible change is the new interface. Instead of foregrounding the file tree and editor, Cursor 2.0 organizes work around agents, allowing several to run in parallel while still letting developers drop back into a traditional IDE view when needed. Combined with git worktrees or remote machines, users can dispatch the same prompt to multiple agents and compare results, a pattern that has been emerging across tools like Claude Code, Devin, and various open-source agent harnesses.

A cluster of supporting features rounds out the release. Browser control is now built into the agent so it can verify its own work by inspecting the DOM and taking screenshots, then iterate. Voice input has been added, and Plan Mode, previously tied to specific models, now works with others including GPT-5. Code review has been reworked to make it easier to scan an agent's diffs across files, addressing a real pain point as agent-generated changesets grow larger. Sandboxed terminals are enabled by default on all platforms, including Windows, and enterprise customers get expanded controls and security options.

The competitive backdrop is significant. GitHub Copilot has been pushing into agent workspaces, Replit and Cognition's Devin emphasize long-running autonomous tasks, and Anthropic's Claude Code targets terminal-native developers. By shipping its own model, Cursor reduces dependence on third-party APIs for both cost and latency, and gains the ability to tune behavior specifically for its agent harness. Whether Composer can hold its own against frontier general-purpose models on harder reasoning tasks remains to be seen, but the strategic logic of vertical integration looks increasingly compelling for IDE vendors.

For teams already invested in Cursor, 2.0 is likely to feel less like an incremental update and more like a re-platforming around agents. Developers who prefer a conventional editor-first experience can still use it that way, but the direction of travel is clear: the IDE is becoming a control surface for orchestrating multiple AI workers rather than a single human typing into a buffer.