On April 2, 2026, Cursor shipped version 3.0. But this isn't an IDE upgrade. They ended the era of AI layered on top of a VS Code fork — and they did it themselves. The brand-new "Agents Window" is now the main interface, and the file editor has become a fallback you summon with Cmd+Shift+P. According to Cursor's own data, a year ago Tab autocomplete users outnumbered agent users 2.5x — now it's flipped, with agent users running 2x ahead. And 35% of internal PRs are now written by autonomous cloud agents. We've crossed from writing code to directing what writes it.
What's Different?
The old Cursor was "AI layered on top of a code editor." The editor was center stage, and AI helped from the side. Cursor 3 flips that assumption. Agents are the unit. Files are secondary. Your sidebar now shows all your agents running side by side — local, cloud, Slack, GitHub, Linear, mobile. Cloud agents even generate demos and screenshots of their output.
The "agent queue" is the new unit of work — not the file editor. The IDE is the fallback you pull up when you need it.
The justification is their own data. CEO Michael Truell breaks AI software development into three eras: ① Tab autocomplete, ② synchronous agents (where you direct each step), ③ autonomous cloud agents. In early 2025, ① dominated. A year later, ③ has overtaken ①. CNBC reported that 35% of merged PRs at Cursor's internal engineering team are now written by autonomous cloud agents.
- $2B ARR milestone
In early 2026, annualized revenue hit $2 billion. Anysphere's valuation doubled to $50B in six months. - Now runs in JetBrains
On March 4, the Agent Client Protocol (ACP) brought Cursor agents into JetBrains IDEs. The "Cursor vs. VS Code" framing is already outdated. - Composer 2 — their own model
Launched March 19. Built on Kimi K2.5 with 4x RL. 200+ tokens/sec, 200K context, 90% cheaper than Claude Opus 4.6 on CursorBench.
The one-liner: this is a pivot to an agent OS. They've laid four foundations at once — a model (Composer 2), an interface (Agents Window), a runtime (cloud VMs), and an enterprise option (self-hosted, launched March 25).
Cursor 3 vs. Cursor 2.x
Here's how the difference plays out in practice, using a backend API feature as the reference point.
| Stage | Cursor 2.x | Cursor 3 |
|---|---|---|
| Setup | Terminal + git checkout, spin up server, open Postman | Run /worktree in Agent Chat — isolated worktree created automatically |
| Implementation | Write controllers, services, and routes by hand | Agent pulls spec via MCP, generates schema-aware code |
| Testing | Run tests in terminal → read the stacktrace → fix in IDE | Agent reads output and auto-fixes failing dependencies via sub-agent |
| Review | Open GitHub, stage → create PR | Stage, commit, PR — all in one flow from the sidebar diff |
The most powerful new feature is /best-of-n. You run something like /best-of-n sonnet, gpt, composer fix the flaky logout test, and multiple models tackle the same task simultaneously, each in an isolated worktree. When they finish, you compare the results and merge the winner. The catch: each model bills separately. Three models = triple the cost.
The local↔cloud handoff works both ways. Start a session locally, push it to the cloud so it keeps running after you close your laptop, then pull it back locally for debugging and push it out again. Long-running tasks don't get cut off. Composer 2 is the default model for this cloud runtime.
"Don't describe it — point at it." Design Mode is genuinely impressive. Hit Cmd+Shift+D to select any UI element directly in the browser, and the agent receives the full HTML, CSS, and bounding box — so "shrink this a bit" actually works.
Community reaction is split. On Hacker News, a recurring criticism went: "Agent-first users want background autonomy; code-first users want precise, synchronous control — you can't satisfy both in one product." Reddit has plenty of "this view disconnects you from the code" pushback. Cost is a real sticking point too — one user spending $2K/week on Cursor reportedly switched to Claude Code Max and got the same productivity for 1/10 the price.
① Plan Mode isn't in the web UI yet (desktop only). ② A bug where agents pause when Agents Window goes inactive was unresolved through early April. ③ Some users saw Composer 2 token consumption spike to ~10x normal due to a cache read issue. ④ Design Mode is Chrome-only (no Firefox or Safari). ⑤ No FedRAMP or HIPAA certification — verify compliance requirements before deploying in regulated environments.
So How Do You Actually Use It?
The interface itself is free to install, so pulling it up costs you nothing. Here's the decision flow:
- You're already a Cursor user
Upgrade → Cmd+Shift+P → "Agents Window." Give it a few days to click. You can still open the old IDE alongside it, so a gradual transition works fine. - You're new to agent workflows
Agents Window is free, but Cloud Agents and Composer 2 start at Pro ($20/month). Start there; heavy users can step up to Pro+ ($60) or Ultra ($200). - You need to compare multiple models
Use /best-of-n in Agent Chat. Worktree isolation means parallel runs don't conflict. Three models = 3x the bill — save it for work that actually matters. - Enterprise with code that can't leave the building
Self-hosted Cloud Agents (Enterprise plan, launched March 25). Code, tool execution, and build artifacts stay on your own infrastructure; workers connect outbound to Cursor cloud via HTTPS.
One more thing — you need to be able to critically read agent-generated code. Multi-file tasks fail regularly. Google DeepMind research on multi-agent systems found error amplification of up to 17.2x in unstructured networks. Even 99% per-agent reliability drops to ~90% after 10 steps in a chain. /best-of-n's worktree isolation helps, but doesn't eliminate this. If you can't read agent output and catch mistakes, your codebase will be in rough shape six months from now.
Truell put it this way: "Cursor is no longer a tool for writing code. It's a tool for building factories that make software." Great as a vision — but in practice, the factory isn't running itself yet. Think of yourself as a director: constantly reviewing and correcting. That said, the interface is clearly built toward that future.
Deep Dive Resources
Cursor official — Meet the new Cursor Design philosophy behind Agents Window, plus primary-source coverage of Composer 2, local-cloud handoff, and the diff-to-PR flow cursor.com
DataCamp — Cursor 3 Feature Breakdown /best-of-n, /worktree, Design Mode, Composer 2 pricing, per-plan feature matrix, and known bugs — all in one place datacamp.com
InfoQ — Why the Community Is Split A roundup of Reddit and Hacker News criticism: the underlying tension between agent-first and code-first models, plus the cost debate infoq.com
Latent Space — Cursor's Third Era The business context: $50B valuation, the Graphite and Autotab acquisitions, and how Cloud Agents overtook the IDE latent.space
