Cursor dropped version 3.0 in April 2026 under the codename Glass. The marketing says it’s the “third era of software development” — fleets of agents, cloud compute, you as the orchestrator.
Four months later, the reality is messier. And more interesting.
What Cursor 3 Actually Is
Cursor 3 isn’t an editor update. It’s a second interface built from scratch — the Agents Window — that sits alongside the classic Cursor IDE. You toggle between them with Cmd+Shift+P → Agents Window.
The Agents Window is a full-screen workspace centered on managing multiple agents in parallel across local, cloud, remote SSH, and mobile environments. The classic editor remains for when you need VS Code extensions, custom themes, or just want to edit files manually.
Cursor’s framing: Three eras of AI coding.
- Era 1 (Tab): Autocomplete, inline suggestions — you write, AI assists
- Era 2 (Agents): Synchronous prompt-response loops — you direct, AI executes
- Era 3 (Glass): Autonomous agent fleets — you define problems, agents ship PRs while you sleep
Thirty-five percent of Cursor’s internal PRs are now created by cloud agents. That’s the signal.
The Headline Features
Agents Window: All Your Agents in One Sidebar
Every agent — local, cloud, kicked off from mobile, web, Slack, GitHub, Linear — appears in a unified sidebar. The interface is multi-workspace by default, meaning a single session can target multiple repositories at once.
You can run dozens of agents simultaneously. Each gets its own context, file access, and Git worktree. They don’t overwrite each other.
Composer 2: Cursor’s Own Frontier Model
Cursor didn’t just build a new UI. They trained their own model. Composer 2 scores 61.3 on CursorBench, 61.7 on Terminal-Bench 2.0, 73.7 on SWE-bench Multilingual — beating Composer 1.5 (44.2/47.9/65.9) by wide margins.
Pricing: $0.50/M input, $2.50/M output. A faster variant at $1.50/$7.50. On paid plans, Composer usage comes from a generous “first-party models pool” — effectively included.
Cloud Agents With Seamless Handoff
Start local → push to cloud when the task gets heavy → pull back to iterate → push again to run overnight. Cloud agents produce demos, screenshots, video recordings, live previews — not just diffs.
The May 13 update added Dockerfile-based environments, build secrets, version history, rollback, and audit logs. Enterprise-ready infrastructure.
/multitask + Worktrees = Parallel Subagents
This is the feature that changes daily workflows. /multitask spawns async subagents that parallelize your request instead of queueing it. Combined with Git worktrees (isolated branches sharing the same .git directory), you can run a fleet of subagents across independent files simultaneously.
Workflow: “Refactor the auth module” → /multitask breaks it into 8 subagents → each gets its own worktree → they run in parallel → you review 8 clean diffs instead of one massive mess.
Cursor Marketplace: Hundreds of Plugins, One Click
MCP servers, skills, subagents — all discoverable and installable inside the Agents Window. Private team marketplaces for internal tooling. This consolidates the fragmented ecosystem of Cursor extensions, MCP servers, and skill files.
Integrated Browser
A built-in browser the agent can drive: open pages, navigate, prompt against local websites. Shortens the frontend feedback loop — the agent sees what it built, iterates, no manual screenshots needed.
The Hard Truth: What’s Broken (Real User Feedback)
The Cursor forum thread “Cursor 3.0 Agentic IDE – Huge Potential, but Blocked by Key Usability Gaps” (821 views, 9 users, detailed breakdown) tells the story early adopters are living.
1. No VS Code Extensions — The #1 Blocker
“No Prettier. No TypeScript tooling extensions. No linting integrations. This makes reviewing and editing agent-generated code significantly harder.”
The Agents Window does not support the VS Code extension ecosystem. No WSL, no Dev Containers, no custom themes. For many developers, this is a non-starter. You have to switch back to classic mode for any extension-dependent work.
2. Diff View Shows Entire Files as “New”
“Diff view shows the entire file as new if touched by an agent, rather than actual changes. Toggle is scoped per file, not global. Forces reliance on Git changes view with infinite scroll.”
Reviewing agent changes — the core workflow — is fundamentally broken in the Agents Window. Users fall back to cmd+e (Git changes) which has its own navigation issues.
3. Fragmented File Navigation
Clicking file links in agent output opens new tabs, pushes you into file browsing mode separate from changes. Constant context switching. No unified “diff-only” ↔ “full file” toggle.
4. File Search Polluted by Compiled Files
No file-type filtering. In TypeScript projects, .js/.d.ts compiled outputs pollute results. cmd+p works better but lacks filtering too.
5. Built-in Browser: No Auto-Refresh, No Session Memory
Changes to local dev apps don’t auto-refresh. No password/session persistence → repeated logins. Undermines the “all-in-one” vision.
6. Missing Keyboard Shortcuts
Close other tabs, close tabs to right, quick jump to browser — all missing. Death by a thousand paper cuts.
When to Use Which Interface
| Task | Use |
|---|---|
| Running 5+ agents in parallel across repos | Agents Window |
| Cloud agents, overnight tasks | Agents Window |
/multitask + worktrees for large refactors |
Agents Window |
| Quick iterations, latency matters | Classic Editor |
| Need Prettier, ESLint, TypeScript extensions | Classic Editor |
| Custom themes, WSL, Dev Containers | Classic Editor |
| Flexible screen splitting, many files open | Classic Editor |
Cursor’s official guidance: The two interfaces are complementary. Most workflows will move between them.
The “Third Era” Bet: Is It Real?
Cursor’s internal data: Agent usage grew 15x in the last year. In March 2025, 2.5x more Tab users than agent users. Now: 2x more agent users than Tab users.
The shift is real. But the Agents Window is early. Think of it like Cursor 1.0 — powerful core, rough edges. The classic editor isn’t going away; it’s the stable fallback.
What Cursor got right: The architecture. Multi-workspace, worktrees, cloud handoff, /multitask, Composer 2, Marketplace — these are the right primitives for agent orchestration.
What they underestimated: How much developers rely on the VS Code extension ecosystem for reviewing agent code. You can’t orchestrate what you can’t properly inspect.
Should You Switch Today?
If you’re already on Cursor Pro/Pro+/Ultra/Teams: Yes, try the Agents Window. It’s a free update. Keep the classic editor open beside it. Use Agents Window for parallel agent work, classic for review and extension-dependent tasks.
If you’re evaluating Cursor for the first time: The Agents Window is the most complete agent-orchestration experience shipping in a commercial IDE today. But factor in the extension gap — if your team lives on Prettier/ESLint/TypeScript tooling, you’ll be toggling constantly.
If you’re on Windows/WSL: The Agents Window doesn’t support WSL yet. That’s a hard blocker for many.
The Bottom Line
Cursor 3 didn’t just ship features. It repositioned the product from “AI-assisted editor” to “agent orchestration platform.” The Agents Window exists because an orchestrator needs a workspace optimized for managing autonomous processes, not editing text.
The usability gaps are real. They’re also exactly the kind of thing that gets fixed fast when you control the full stack (Cursor forked VS Code precisely to avoid extension API limitations).
Cursor’s bet: The third era arrives in months, not years. The Agents Window is where they’re building it.
My bet: By Q4 2026, the extension gap closes (they’re already tracking it), the diff view gets fixed, and the Agents Window becomes the default for anyone doing serious agent work. The classic editor becomes “legacy mode” — still there, still supported, but not where the future happens.
Quick Start: Your First Hour in Cursor 3
# 1. Update to latest
# Cursor auto-updates, but check Help > About > version ≥ 3.0
# 2. Open Agents Window
Cmd+Shift+P → "Agents Window: Open"
# 3. Try /multitask on a real task
# In an agent chat: "/multitask refactor the user service into separate modules"
# 4. Spin up a cloud agent
# Click the cloud icon in sidebar → "New Cloud Agent" → watch it run while you close laptop
# 5. Browse Marketplace
# Sidebar → Marketplace → install an MCP server or skill
# 6. When extensions block you
Cmd+Shift+P → "Open Editor Window" → classic Cursor with full VS Code extension supportRelated Reading
- Cursor vs AmpCode: IDE Daily Driver or Terminal Agent?
- Cursor vs GitHub Copilot CLI: IDE vs Terminal
- Best Coding Agents 2026 Decision Guide
- Claude Code vs Cursor 2026 Comparison
Related articles
- Claude Code Alternatives in 2026: 12 Options Compared
- Cursor Background Agents Changed How I Code
- Why I Stopped Using Copilot and Went Full Terminal Agent
The agent wars aren’t about prompts anymore. They’re about what sticks after the prompt ends. Cursor 3 just made orchestration stick — even if the diff view still needs work.
Comparing agents is half the work. aiFiesta gives you every major AI model in one chat for $12/mo. Test workflows side-by-side.