Cursor 3 shipped April 2 and the framing is deliberately different from the prior releases: not a better AI editor, but a workspace for managing fleets of agents.
The previous two versions of Cursor followed the same basic model as every AI-assisted IDE: one agent, one conversation, one task at a time. You open a tab, describe a change, review a diff, repeat. The agent lives inside the editor. Cursor 3 inverts that relationship. The editor becomes a viewport into a system where many agents may be running simultaneously — some on your local machine, some in the cloud — across multiple repositories.
What actually changed
The centerpiece is the new sidebar that surfaces all active agents regardless of where they were started. An agent kicked off from Slack, from GitHub, from the Cursor mobile app, or from Linear shows up alongside the one you just started locally, in the same panel. Cloud agents produce screenshots and demos of their work so you can verify what they’ve done before pulling changes down.
The cloud/local handoff is practical: you can spin up a long-running agent, close your laptop, and pick it up later — either continuing in the cloud or moving it back local for testing. This matters for the kinds of tasks that take twenty minutes to run and die if you lose your network connection.
Cursor is also introducing Composer 2, described as an in-house frontier coding model with high usage limits, available within local agent sessions. They don’t publish its training details, but the move signals Cursor is no longer content to be purely a wrapper around third-party models. Having your own model means direct control over context management, cost, and the specific failure modes that matter for coding workflows.
The Cursor Marketplace is the third big piece: a browsable catalog of plugins that extend agents with MCPs, skills, and subagents. Teams can set up private marketplaces. This is a platform bet — Cursor positioning itself as the deployment target for tooling the community builds, not just a product.
Lesser features worth noting: full LSP support (go-to-definition, type checking in the new diffs view), an integrated browser the agent can navigate, and a streamlined staging/commit/PR flow that handles the full cycle without leaving the tool.
What this is really about
The language Cursor uses in the announcement is worth reading carefully: “Engineers are still micromanaging individual agents, trying to keep track of different conversations, and jumping between multiple terminals, tools, and windows.” That’s an accurate description of how most people use AI coding tools today, including current Cursor. The problem they’re solving is the coordination overhead of agentic workflows, which has grown as agents have gotten more capable and the tasks have grown longer.
The model being built toward is agents that run more autonomously and surface results at a higher level of abstraction. Whether Composer 2 and the new UX are enough to get there is still an open question — agent reliability at the long-horizon tasks where coordination overhead matters most is still uneven. But the direction is clear: the IDE is becoming the shell, not the primary interface.
Cursor 3 is available now via the standard update channel; the multi-agent interface is behind Cmd+Shift+P → Agents Window.