Anysphere recently released Cursor 3, a redesigned interface for its AI coding tool that shifts the primary interaction model from file editing to managing parallel coding agents. The new interface was built from scratch rather than extending Cursor's existing VS Code fork, though users can switch back to the full IDE at any time.
The central idea behind Cursor 3 is that developers spend more time orchestrating agents than editing code directly, and the tooling should reflect this. Co-founders Michael Truell and Sualeh Asif describe it as "a unified workspace for building software with agents," framing it as a step toward what Truell has called the third era of software development, where fleets of agents work autonomously to ship changes. Cursor's own usage data supports the shift: as recently as March 2025, 2.5x as many users were using tab completion as agents. That ratio has now completely inverted, with twice as many users running autonomous agents. Internally, 35% of merged pull requests at Cursor's own engineering team are written by autonomous cloud agents.
The interface surfaces all running agents, both local and cloud, in a single sidebar. Agents kicked off from mobile, web, desktop, Slack, GitHub, or Linear all appear in one place. Cloud agents generate demos and screenshots of their work for human review. Users can run multiple agents in parallel across different repositories, which the previous interface didn't support natively.
(Source: Cursor blog post)
The local-to-cloud handoff is worth noting for teams evaluating these tools. An agent session running locally can be moved to the cloud to continue working while the developer is offline or is working on another task. In the other direction, a cloud session can be pulled back to local for hands-on editing and testing. Cursor says the handoff is fast, though they don't publish latency numbers. The cloud execution uses Composer 2, Cursor's own frontier coding model, which they say comes with higher usage limits than third-party models.
A new plugin marketplace lets teams extend agents with MCPs, skills, and subagents. Organizations can also set up private team marketplaces for internal plugins, which addresses a governance need similar to what AWS is tackling with its Agent Registry at a different layer of the stack.
Community reactions on Reddit and Hacker News have been sharply divided, with much of the pushback centered on whether Cursor is abandoning the IDE-first identity that attracted its user base.
As one Reddit user put it:
This view makes you lose any connection to your code... I specifically stay with Cursor because it's so good at being an IDE.
While one commenter on Hacker News articulated why the agent-first model feels wrong for some workflows:
Reviewing and testing code, constantly switching contexts, juggling model contexts, coming up with prompt incantations to coax the model into the right direction... is so mentally taxing and full of interruptions that it's practically impossible to achieve any sort of flow state. Working with LLMs is the complete opposite of this.
Another Hacker News commenter framed the underlying product design tension:
Agent-first needs ambient, background autonomy. Code-first needs precise, synchronous control. Trying to do both in one product means you're always making tradeoffs that frustrate one half of your users.
In addition, users in the Reddit thread raised concerns about vendor lock-in. One commenter argued that "the proper agent command center I would want to use is the one that I could manage all AI agents I have, not lock into one vendor." Lee Robinson, a Cursor moderator, responded that Cursor supports models from all vendors.
Furthermore, the cost question recurred in both threads. One Hacker News commenter who switched from Cursor to Claude Code reported spending "$2k a week with premium models" on Cursor before switching to Claude Code Max, where they described being "equally as prolific and paying 1/10th the price." A Reddit user claimed the same model consumed dramatically different token budgets depending on the harness, estimating 12% daily usage for a workflow in Claude Code versus 80% for the same workflow in Cursor.
If these reports are representative, they suggest that how a harness constructs context and manages tool calls can have a significant impact on effective cost, a consideration that matters as much as the interface design for teams evaluating agent-first tools.
The release positions Cursor more directly against Claude Code and GitHub Copilot's agent mode, both of which have also been moving toward autonomous multi-step execution. The difference in approach is instructive: Claude Code operates as a CLI tool that developers invoke from the terminal, GitHub Copilot remains embedded in the IDE, and Cursor is now building a purpose-built surface that treats the IDE as a fallback rather than the primary interface.
Cursor 3 is available now. Users can access the new interface by upgrading Cursor and running Cmd+Shift+P -> Agents Window.