2.0
CursorSummary
This release introduces significant new capabilities, including the Multi-Agents editor for running up to eight agents in parallel and the new, faster Composer coding model. Key stability improvements include Sandboxed Terminals going GA on macOS and major performance boosts for Language Server Protocols. Users can now also control the agent via Voice Mode and define centralized Team Commands.
New Features
- Introduced a new Multi-Agents editor featuring a sidebar for managing agents and plans.
- Enabled running up to eight agents in parallel on a single prompt using isolated environments (git worktrees or remote machines).
- Introduced Composer, a new agentic coding model that is 4x faster than similarly intelligent models.
- Launched Browser functionality to General Availability (GA) with added support for Enterprise teams.
- Enabled Browser to be embedded in-editor, including new tools for element selection and forwarding DOM information to the agent.
- Enabled defining custom commands and rules for your Team via the Cursor dashboard, automatically applied to all members.
- Introduced Voice Mode, allowing control of the Agent using built-in speech-to-text conversion.
- Enabled defining custom submit keywords in settings to trigger the agent to begin running in Voice Mode.
- Enabled creating plans with one model and building the plan with another, with options for foreground, background, or parallel agent execution.
- Enabled sharing custom rules, commands, and prompts with the entire team via deeplinks.
- Enabled Enterprise teams to distribute hooks directly from the web dashboard, including managing drafts and OS targeting.
- Enabled viewing a timestamped Audit Log of admin events including user access, setting changes, and team rule edits.
Bug Fixes
- Fixed file conflicts when running multiple agents in parallel by using isolated environments.
- Fixed memory leaks and improved overall memory usage across the application.
- Improved quality when working with Agent across all models, especially GPT-5 Codex, due to harness improvements.
Improvements
- Improved Code Review experience to easily view all changes from the Agent across multiple files without needing to jump between files.
- Sandboxed Terminals are now GA for macOS, running agent commands securely by default with read/write access to workspace but no internet access (unless allowlisted).
- Improved performance of loading and using Language Server Protocols (LSPs) for language-specific features, especially noticeable when working with agents and viewing diffs.
- Made Python and TypeScript LSPs faster by default for large projects with dynamically configured higher memory limits.
- Improved the prompt UI by showing files and directories inline as pills.
- Improved copy/pasting prompts with tagged context.
- Removed many explicit context menu items (@Definitions, @Web, @Link, @Recent Changes, @Linter Errors) as the Agent can now self-gather context.
- Improved the underlying harness for working with Agent across all models.
- Improved the experience of sending agents to the cloud from the editor.
- Enabled Enterprise admins to enforce standard settings for Sandboxed Terminals (sandbox availability, git access, network access) across their team.