Table of Contents
- What Cursor 3 Glass Actually Replaced
- The Five Workflow Changes That Matter
- A Real Example: Fixing a UI Bug End to End
- Cursor 3 vs Claude Code vs Codex
- Who This Is For (and Who It Isn’t)
- The Pricing Math
- When Not to Use Cursor 3
- Do This First
- What This Signals for the Rest of 2026
- FAQ
Cursor 3 didn’t ship a new feature on April 2, 2026. It shipped a new identity. The team behind the most-installed AI code editor of the past two years rebuilt the product from scratch under the codename Glass and pushed it live with one structural change at the center: the file tree is no longer the default surface. An Agents Window is.
If you’ve spent any time on developer Twitter since launch day, you’ve seen the screenshots. Three or four agents running in parallel, each working in its own git worktree, each producing diffs the developer compares before merging. That used to be a stretch goal. Now it’s the front door of the product.
Cursor’s own framing is direct. “Cursor 3 is a unified workspace for building software with agents,” the team wrote in their launch post. The version of you that opened a file and started typing is being designed out of the product. The version that dispatches work to agents and reviews output is being designed in. Here’s what actually changes day to day, where it breaks, and how to use it.
What Cursor 3 Glass Actually Replaced
The Composer pane is gone. In its place is the Agents Window, launched with Cmd+Shift+P → Agents Window, which is now the default surface when you open Cursor 3. From there you spawn agents that run in any of four environments: locally, in a separate git worktree, in the cloud, or over remote SSH. You can move an agent from one environment to another mid-task. You can run many in parallel. The IDE — the file tree, the editor pane, the terminal — still exists. It is just no longer where the work starts.
This is not a UI refresh. It is an admission that engineers in 2026 spend their time managing fleets of agents, not typing into editors. Cursor’s own framing calls this “the third era of software development, where fleets of agents work autonomously.” Whether or not you buy the era framing, the product now assumes you do, and the keyboard shortcuts have been redesigned around that assumption.
The Five Workflow Changes That Matter
1. Agent Tabs in the editor. Multiple chats side by side or in a grid. You can have one agent fixing a failing test, another writing a database migration, a third triaging an issue queue. Switch between them like browser tabs. Each tab keeps its own context, history, and selected model.
2. Design Mode in the in-app browser. Hit Cmd+Shift+D to toggle Design Mode on. Shift+drag to select a region of your running app. Cmd+L sends the selected element straight into an agent’s chat. Alt+click does the same to the input field. The agent now knows exactly which div, button, or component you’re talking about. UI bug reports stop being “the form looks weird on mobile.”
3. The /worktree command. Type /worktree in the agent input and Cursor creates an isolated git worktree for the task. The agent’s changes happen there. Your main branch stays untouched. When you’re satisfied, you merge. When you’re not, you delete the worktree and the agent’s work is gone with no impact on your repo state. This single primitive removes the biggest reason developers were nervous about letting agents write code: the fear of polluting the working tree.
4. The /best-of-n command. Runs the same prompt across multiple models in parallel, each in its own isolated worktree. You compare outputs side by side and pick the one to merge. For non-trivial refactors, this is the workflow that saves the most time. Sonnet may write the cleanest version. GPT-5.5 may catch an edge case Sonnet missed. Composer 2 (Cursor’s in-house frontier coding model) may handle the project’s idioms best. /best-of-n lets you stop guessing.
5. The Await tool. Agents can now wait for background shell commands or subagents to finish, or for specific output strings like “Ready” or “Error.” This sounds small. It is the difference between an agent that gets stuck on a long-running build and one that resumes intelligently the moment the build finishes. Combined with cloud agents, it means you can dispatch a multi-step task at the start of a meeting and come back to a finished diff.
A Real Example: Fixing a UI Bug End to End
Open Cursor 3. Press Cmd+Shift+P → Agents Window. Open the running app in your in-app browser. Hit Cmd+Shift+D to enter Design Mode. Shift+drag over the broken modal. Press Cmd+L to send it to the agent.
Type into the agent input: /worktree fix the modal so the close button is keyboard-accessible and the focus trap works on tab.
The agent spawns a fresh worktree, makes the changes, runs your test suite, waits for the result via the Await tool, and reports back with a diff. You review it. If you like it, merge. If not, you spawn another agent with /best-of-n and let three models propose alternatives in parallel.
This used to take ten or fifteen minutes of context switching: clicking around the editor, dragging snippets into a chat, copy-pasting suggestions back, manually verifying tests. Now it is a single chain of commands without leaving the Agents Window. The time savings are real but secondary. The bigger shift is psychological: you stop being the typist and start being the reviewer.
Cursor 3 vs Claude Code vs Codex
Cursor’s positioning is implicit. The launch post does not name competitors. But the product unmistakably answers the same problem Claude Code and OpenAI’s recently overhauled Codex desktop app are answering: how should a developer dispatch and review work from agents in 2026?
The three converge on the same idea — agents as the default surface — but they make different bets:
Cursor 3 Glass bets on a multi-environment Agents Window with worktrees as the unit of safety. Best for developers who already use Cursor and want zero ramp time on the model selection layer.
Claude Code bets on terminal-native agents with deep file system access and skills as reusable prompts. Best for developers comfortable in the CLI who want minimal UI overhead.
OpenAI Codex (April 16 update) bets on background computer use, where Codex agents see, click, and type with their own cursor across all your Mac apps. Best for developers whose work spans tools beyond the editor.
None of these is a strict superset of the others. Our coding tools comparison covers the head-to-head detail.
Who This Is For (and Who It Isn’t)
If you’re a solo developer or indie hacker shipping fast, Cursor 3 is the largest single workflow improvement of the year. Parallel agents plus /worktree means you can attempt three risky changes at once without polluting your repo. That is a posture difference, not a productivity bump.
If you’ve spent two years building muscle memory around the Composer pane, the first 48 hours will hurt. The keyboard shortcuts changed. The Agents Window puts decisions in front of you that the Composer pane handled implicitly. Some users on the Cursor forum reported turning off the new interface and reverting to the file-tree default for the first week. That is an option, not a defeat.
If you’re already deep in Claude Code or running parallel agents through OpenClaw or Claude Cowork, Cursor 3 does not replace them. It joins them. Many developers will end up running Cursor 3 inside the editor and Claude Code in a terminal split for different categories of work.
If you’re a VS Code purist or a vim diehard who wanted Cursor only for inline completions, Cursor 3 is going to feel like Cursor abandoned you. It didn’t, but the default surface is no longer for you. The completions still work. You just have to ignore the Agents Window’s gravitational pull.
The Pricing Math
Cursor 3 ships across the existing tier structure: free Hobby, Pro at $20/month, Pro+ at $60, Ultra at $200, and Teams at $40 per user per month. Annual billing saves 20%.
The credit math is the part that bites. Auto mode is unlimited and does not draw from your credit pool. The moment you switch to a premium model — Claude Sonnet, GPT-5.5, or Composer 2 — you start spending. A $20 Pro pool covers roughly 200 to 250 Sonnet requests in practice, depending on token volume per request.
/best-of-n is incredible. /best-of-n is also the fastest way to drain a Pro plan, because it bills for every model you ran in parallel. If you plan to use it on every task, budget Pro+ at minimum. The Ultra tier is overkill for solo developers and the right floor for small teams that share a single account.
When Not to Use Cursor 3
If your repo is small, your tasks are well defined, and you mostly want completions and quick edits, the Agents Window is overhead. Stay on Auto and live in the editor pane. Cursor 3 does not make Cursor 2’s workflow worse. It just makes a heavier workflow available.
If you work primarily in regulated environments where every change has to flow through a code review system tied to specific people, /best-of-n introduces audit complexity that may not be worth the time savings. Solo Pro use is fine. Enterprise use needs a policy conversation first.
If your work is mostly exploratory data analysis in notebooks, Cursor 3’s strengths don’t apply. Use a notebook-native agent surface. The Agents Window assumes a long-lived repo and a test suite that can validate changes.
Do This First
Update Cursor. Press Cmd+Shift+P. Type “Agents Window” and hit return. Spawn an agent. Type /worktree run the test suite and fix the most recently failing test. Watch the agent work in an isolated worktree while your main branch sits untouched. That single 30-second loop is the entire pitch.
If it lands, the rest of the workflow stack — Design Mode, Agent Tabs, /best-of-n — unlocks naturally. If it doesn’t, the older Composer-style flow is still there and your monthly invoice doesn’t change.
What This Signals for the Rest of 2026
The IDE is being demoted everywhere, not just at Cursor. Claude Code, OpenAI’s Codex desktop app, and GitHub Copilot in Visual Studio have all moved their default flows toward agent management surfaces in the last 60 days. Cursor 3 is the most aggressive expression of the trend, but it isn’t alone. By the end of Q3 2026, expect every major coding tool to default to an agent-first surface and treat the file tree as a power-user fallback. The broader agent stack story tracks the same convergence at a higher altitude.
The interesting question is no longer whether the agent-first model wins. It already won. The interesting question is whether the next generation of developers — the ones who started learning to ship code in 2025 — ever uses a file tree as their primary surface at all. If they don’t, the IDE as a category becomes what command-line FTP clients are today: still installed, occasionally useful, never the first thing you open.
FAQ
Is Cursor 3 a free upgrade?
Yes. Cursor 3 ships across all existing tiers including the free Hobby plan. There is no separate “Glass” upgrade. Open the app and update.
Do my old Composer chats migrate to the Agents Window?
Cursor 3 imports your prior chat history into the Agents Window on first launch. Existing chats appear as agent tabs you can resume or close. The migration is one-way; downgrading to Cursor 2 means losing the new tab metadata.
Can I disable the Agents Window and go back to the Composer pane?
Yes. The Composer pane is still available behind a setting. Some users on the official Cursor forum recommended this for the first week while building muscle memory for the new shortcuts. It is a supported configuration.
How does /worktree affect my git history?
A git worktree is a standard git feature that creates a separate working directory tied to a different branch. The agent works there in full isolation. When you merge, you merge a clean branch into your main one. There is no impact on history beyond the merge commit you choose to create.
Does /best-of-n let me pick which models to compare?
Yes. By default Cursor proposes the strongest available models on your tier (typically Sonnet, GPT-5.5, and Composer 2 on Pro). You can override the model list in the command flags. Each model runs in its own worktree so you can diff outputs cleanly.
How does Cursor 3 compare to Claude Code for terminal-heavy workflows?
Claude Code is purpose-built for terminal-native flows and remains the better choice if you do most of your work in a CLI. Cursor 3 is the better choice if you want a graphical Agents Window, in-app browser, and worktree primitives without leaving the editor. Many developers run both for different categories of work.