Summary: Cursor 3, released on April 2, is an agent-centered workspace rather than a conventional IDE, with multi-repo support and one sidebar for local and cloud agents. It adds bidirectional local-to-cloud handoff, diffs and PR management, a built-in browser, a plugin marketplace, and optional private team marketplaces. The article notes that the launch includes no pricing or benchmark numbers, so teams should test the mechanisms directly.
Writing software no longer means typing most of the code yourself. Over the past year, daily work has shifted toward teaming up with agents that produce the bulk of the code. Cursor has a name for where this leads — the third era of software development, in which fleets of agents ship improvements on their own. The destination is easy to picture, yet the same announcement concedes how far current practice sits from it: engineers still micromanage one agent at a time, hopping across conversations, terminals, tools, and windows while trying to keep track of where everything happened.
Cursor 3, which landed on April 2, is the company’s reply to that distance: a unified workspace for building software with agents. In its own telling it runs faster, looks cleaner, and does more, adding a multi-repo layout, seamless handoff between local and cloud agents, and the freedom to step back into the Cursor IDE whenever needed. What follows unpacks each mechanism — what it actually changes, what the launch leaves unsaid, and the trade-offs worth weighing before adoption.
Not a Fork Anymore: Starting From a Blank Page
How Cursor began is itself a clue: rather than ship a VS Code extension, the team forked the editor outright, because owning the interface was the entire point. Cursor 3 carries that logic further — the new interface is not painted onto the old IDE; it starts from zero with agents at the center.
The weight of that choice exceeds its wording. Where an IDE revolves around files, an agent amounts to a visitor parked in the sidebar; where a workspace revolves around agents, browsing files, reading diffs, and driving a browser become the instruments of verification and deeper inquiry. The route back to the Cursor IDE stays open, and the company frames this as gradual migration, never a forced switch.
One Sidebar, Every Agent
Multi-workspace support comes built into the new interface, so humans and agents can operate across different repos. Every local and cloud agent shows up in that single sidebar, no matter whether the task began on mobile, web, desktop, Slack, GitHub, or Linear. A cloud agent hands over demos and screenshots when it finishes, so verifying results no longer means trawling logs. Per the company, this is the same experience as cursor.com/agents, now brought into the desktop app.
Descend a level and the change is about lifting where developers stand: the status and output of each task greet you first, and files or terminals are there for when you truly need to go deeper. With several agents in flight, what gets saved is the cost of switching context — very often where an engineer’s hours actually vanish.
Handoff That Works Both Ways: Local Meets Cloud
Of everything shipped here, handoff carries the most hands-on value, and it runs in both directions:
| Way of moving | Best when |
|---|---|
| Cloud to local | You need to edit the code yourself, or bring a service up on your own machine for interactive testing |
| Local to cloud | The job is long-running and has to keep going after the laptop closes, or you would rather get started on the next thing |
Per the company, moving an agent across environments is “really fast.” On the model side, the partner is Composer 2, Cursor’s homegrown frontier coding model, pitched by the vendor as carrying high usage limits and suiting rapid iteration. What handoff removes is not a verdict on “which environment is stronger” — it is the need to explain the work over again each time the environment changes. For long-running jobs, that transfer cost is regularly harder to manage than model inference time.
Commit Through Merged PR, With IDE Strengths Untouched
The rebuilt diffs view arrives with a simpler UI, so editing and reviewing changes take less effort; when you are satisfied, staging, committing, and managing PRs happen right there. The code-comprehension side lost nothing: files can be inspected, go to definition still works inside the editor, and full LSPs are on board. A built-in browser completes the set — it opens local websites, moves around them, and accepts prompts against them, which proves its worth most when checking front-end work.
Alpha users, in feedback the company quotes, single out how Cursor 3 gathers the IDE’s strongest parts and its newer agent-first capabilities into one interface. This is qualitative feedback, yet it matches the direction the mechanisms are pointed.
A Plugin Marketplace, Plus a Private One for Teams
Hundreds of plugins populate the Marketplace, extending agents by way of MCPs, skills, subagents, and other means, each a single click to install; a team can additionally raise its own private team marketplace. The upshot: agent tooling stops being a personal config file and becomes an asset the whole team maintains — and for a team rolling agent workflows out, that governance entry point outweighs how many plugins there are.
A Launch That Brought No Numbers
It deserves an honest flag: nowhere in the announcement is there pricing, a percentage, or a benchmark figure — Composer 2’s usage limits stop at the qualitative “high.” Judging a launch shaped like this means leaning on the verifiability of the mechanisms themselves — does the sidebar truly pull in tasks from every origin, does handoff truly carry over without context being rebuilt — not on quantified assurances from the vendor. The third-era picture of fleets of agents shipping autonomously is the company’s stated position; the adoption call should rest on mechanisms you can operate today.
For Builders: Getting Underway
Turning it on takes one step: upgrade Cursor and run Cmd+Shift+P -> Agents Window. On pacing the adoption, three suggestions. One, get to know the sidebar and handoff on a single task before letting agents run in parallel — a higher parallel count does not by itself raise output; task boundaries and acceptance criteria are what do. Two, send your longest task to the cloud ahead of everything else: it keeps running through a closed laptop, the easiest benefit to verify and the fastest to feel. Three, if the team has already built up internal MCPs or skills, open a team marketplace without waiting, and put plugin governance in front of the workflow rollout.
The company says plainly that the interface will not stop here: stronger models will unlock interaction patterns that do not exist yet. For builders, what makes Cursor 3 worth serious attention is not the length of its feature list — it is that multi-agent collaboration has been pulled out of the realm of vision and onto a workbench you can operate today.