Changelog 3.9 · Aug 5, 2026
Orchestrate multiple Agents across local, worktree, cloud, and SSH in parallel; instantly return to full IDE depth when needed. Cursor 3.x brings Multi-agent Swarm, Deep Research Mode, Sidecar Preview, Auto-review, Composer 3.0, and more.
Fonts: Inter, JetBrains Mono (Bunny Fonts · SIL OFL)
Cursor 3.x
New Agent-first interface supporting parallel execution across local, worktree, cloud, and SSH. Unified sidebar manages all Agents, including those from mobile, web, Slack, GitHub, Linear, Jira, and Microsoft Teams. Cloud Agents auto-generate demos and screenshots for verification.
Annotate and select UI elements directly in the built-in browser for precise Agent feedback. Keyboard shortcuts: ⌘⇧D to toggle, ⇧+drag to select, ⌘L to add to chat.
New in 3.6: Run mode that reduces approval fatigue. Allowlisted commands run immediately, sandboxed tasks execute safely, other requests routed to a classifier subagent for autonomous decision-making.
/loopNew in 3.5: Run a prompt repeatedly on a schedule until a goal is met or stopped. Great for monitoring deploy status, continuous testing, and long-running tasks.
New in 3.6: Major intelligence upgrade. Better at sustained work on long-running tasks, follows complex instructions more reliably, smoother collaboration experience.
Tools return richer structured data. Marketplace offers hundreds of plugins (MCP, Skills, Subagents, Rules, Hooks) with one-click install. Teams can create private marketplaces with distribution policies.
3.1 → 3.9 Updates
Cursor 3.0 has been continuously iterating, bringing Multi-agent Swarm, Deep Research Mode, Sidecar Preview, Auto-review, and more.
Swarm-style multi-agent collaboration: architects, coders, and testers work in parallel on medium to large projects with automatic progress consolidation.
Learn more →Agents interact directly with built-in headless/desktop sandboxed browsers to click UI elements, analyze screenshots, and intercept network traffic for automated Web testing.
Sidebar rendering of live HTML/React components, DOM tree inspection, and responsive viewport toggles to streamline frontend development.
Intelligently captures shell errors and missing dependencies, auto-completing parameters or correcting syntax to reduce terminal friction.
Autonomously inspects multi-source docs, audits codebase dependencies, and simulates architecture refactoring to produce high-precision technical plans.
Monitors unit test and linter failures, forming an autonomous loop of hypothesis, code edit, and re-testing until all assertions pass.
New run mode that reduces approval fatigue. Allowlisted commands execute immediately, sandboxed tasks run safely, other requests routed to a classifier subagent that decides: allow, try different approach, or ask for approval. Custom instructions supported.
Learn more →Major intelligence upgrade: better sustained work on long tasks, more reliable complex instruction following, smoother collaboration. New faster tier, double usage for first week.
Share interactive canvases (reports, dashboards, custom interfaces) created by Agents with your team. Recipients can open live snapshots in the browser without logging into Cursor. View team's shared canvases in Dashboard.
/loop SkillRun a prompt repeatedly on a local schedule until a goal is achieved or stopped. If no fixed interval specified, the agent decides when to wake. Examples: "check deploy status every 5 minutes", "work on feature until tests pass".
Automations now available in Agents Window. Support multi-repo automations (reason across codebases) and no-repo automations (monitor tool signals). 5 new no-repo templates: Slack digest, product analytics, FAQ, finance, customer health.
Assign work items to Cursor in Jira, or @Cursor in comments to kick off cloud agents. Microsoft Teams support lets you @Cursor to delegate tasks or pull information into Teams.
Single cloud Agent environment can include multiple repos, config reused across sessions. Agents reason about cross-repo impacts. New Dockerfile build secrets, 70% faster rebuilds via layer caching, agent-led credential validation.
Three tabs: Reviews (inline threads and top-level comments), Commits (commit history), Changes (file tree and changes picker). End-to-end from creation to merge in one place.
Identify independent parts of your plan and run them simultaneously using async subagents. Dependent steps kept in order. Explore subagent supports model selector for cheaper "find relevant code" tasks.
Analyzes chat transcript, groups commits by logical concern, opens one PR per group. Detects real dependencies and adds "depends on" notes. No more giant PRs with unrelated changes.
Core Model
Cursor's latest in-house coding model, built on Kimi K2.5 with massive RL training. Matches Claude Opus 4.7 and GPT-5.5 on benchmarks at 1/10 to 1/60 the cost.
| Benchmark | Composer 2.5 | Opus 4.7 | GPT-5.5 | Composer 2 |
|---|---|---|---|---|
| SWE-Bench Multilingual | 79.8% | 80.5% | 77.8% | 73.7% |
| Terminal-Bench 2.0 | 69.3% | 69.4% | 82.7% | 61.7% |
| CursorBench v3.1 | 63.2% | 61.6% | 59.2% | 52.2% |
Standard: $0.50/$2.50/M tokens (~$0.07/task). Fast: $3/$15/M tokens (~$0.44/task). Opus 4.7 ~$4.10, GPT-5.5 ~$4.82 per task.
Injects correction hints at error points during training. 25x more synthetic tasks including "feature deletion" scenarios.
30% faster than Standard, avg 6.7 min/task. 6x cost. Default in Cursor.
Learn & Demo
Curated YouTube tutorials from the Cursor team and trusted creators. For the most authoritative product definitions and dates, check the official links below.
Authoritative docs for features covered on this page.
Version Evolution
Cursor chose to fork VS Code rather than just making a plugin, enabling deep transformation of interaction and model pipelines.
Up to 8 Agents in parallel, with Composer, Browser GA, Sandboxed Terminals, Plan Mode, and Voice Mode.
Natively multi-workspace / multi-repo, desktop and cloud Agent experience unified. Design Mode brings browser annotations into Agent.
Build in Parallel, Split PRs, Unified PR Review. Agents work on independent steps simultaneously.
Multi-repo cloud environments, Dockerfile build secrets, 70% faster rebuilds. Shared Canvases, /loop, multi-repo/no-repo automations. Jira & Teams integration.
Auto-review reduces approval fatigue with tiered execution. Composer 2.5 major intelligence upgrade for long tasks and complex instructions.
Introduces Deep Research Mode (multi-source doc/codebase audit) and Autonomous Debugger (auto-capturing test/lint errors with fix cycles).
Workspace Sidecar live sidebar component rendering & DOM debugging. Terminal Auto-Correction syntax auto-completion for shell errors.
Multi-agent Swarm enables swarm-style Agent collaboration across architecture, coding, and testing; Sandboxed Browser Action for automated UI testing.
Pricing
Bugbot code review: now usage-based billing (Teams/Individual). Configurable effort levels (Default/High/Custom). View full pricing →
Competitors
| Dimension | Cursor | Windsurf | Copilot | Claude Code | Augment |
|---|---|---|---|---|---|
| Main Interface | Standalone AI IDE | Standalone AI IDE | Multi-editor plugin | Terminal Agent | IDE Extension |
| Proprietary Model | Composer 3.0 | SWE-1.6 | None | None | Context Engine |
| Multi-Agent | Swarm | ||||
| Deep Research Mode | |||||
| Sidecar Web Preview | |||||
| Cloud Agent | |||||
| Large Codebase Understanding | |||||
| Auto-review / Reduce approvals | |||||
| Free Tier | Limited | Most generous | Limited | Pay per token | None |
| Privacy / Self-host | SOC 2 | ||||
| Starting Price | Free / $20 | Free / $15 | Free / $10 | Pay per token | $30 |