Directing a team of agents, not just prompting a single model.
A developer sits down to build a new app feature. Instead of writing the first line of code, they type a plain English description into a chat window, pick a language model, and watch the software take shape — complete with a demo video showing exactly what was built. That's the pitch behind Cursor 3, the latest release from Anysphere Inc., the startup behind one of the more closely watched AI coding tools in the industry.
Cursor has been building toward this kind of workflow for a while. The company has pulled in more than three billion dollars in funding, with Nvidia and Google among its backers, and its editor already handled tasks like generating new application features and hunting down bugs. The engine underneath draws on both commercial models like Anthropic's Claude and algorithms Cursor has developed on its own. Version 3 doesn't abandon that foundation — it builds a more structured architecture on top of it.
The centerpiece of the new release is a chatbot interface that routes work through multiple AI agents simultaneously. Some of those agents live in the cloud, where they have access to more computing power and can run in parallel — useful when a task is large enough that waiting on a single process would slow everything down. Others run directly on a developer's machine, which is slower but gives the developer hands-on access: they can open the generated code, edit it manually, and run tests without anything leaving their local environment.
The two modes aren't mutually exclusive. Cursor 3 lets developers move work between them — spinning up a fleet of cloud agents to generate a block of code, then handing it off to a desktop agent for local review and refinement. A new sidebar serves as the control panel for all of this, giving developers a single place to see and manage both varieties of agent regardless of where they're running.
Cursor is also pointing developers toward Composer 2, an in-house language model it released last month, as a particularly good fit for this kind of hybrid workflow. The company describes it as more cost-efficient than several of the other models the platform supports, which matters when you're running multiple agents in parallel and the compute costs can add up quickly.
For developers working on user interfaces, a new Design Mode changes how edits get made. Rather than hunting through code to adjust a button's behavior or a layout's spacing, a developer can click directly on an interface element, type a description of what they want changed, and let the agents handle the implementation. It's a small thing in isolation, but it's the kind of friction reduction that compounds over a long workday.
Transparency is woven into how the platform reports its own work. Every task Cursor completes comes with a step-by-step breakdown — plain language explanations of each sub-step, flagged errors, and screenshots of the output at various stages. Developers can steer the process by typing feedback at any point, which keeps a human in the loop without requiring them to drop into the code directly.
A handful of smaller additions round out the release. One new shortcut lets developers send the same request to several language models at once and choose whichever response looks most useful. Another improvement tightens up the process of reviewing code changes before they go to production — a step that tends to be tedious and error-prone, and one where a little automation can prevent a lot of headaches.
The broader question hanging over all of this is whether tools like Cursor 3 change not just how fast developers work, but what kind of work they spend their time on. As the agents handle more of the mechanical labor, the role of the developer shifts toward directing, reviewing, and deciding — which is either liberating or unsettling depending on who you ask. How that plays out in real teams, over real projects, is the thing worth watching.
Notable Quotes
Composer 2 is particularly well suited for hybrid cloud-and-desktop workflows and is more cost-efficient than several other supported models.— Cursor (Anysphere Inc.), per company statement
The Hearth Conversation Another angle on the story
What's actually new here — isn't Cursor already an AI coding tool?
It was, but version 3 is less about the editor and more about orchestration. You're now directing a team of agents, not just prompting a single model.
What's the difference between the cloud agents and the desktop ones?
Power versus control. Cloud agents can run many tasks at once and finish faster. Desktop agents are slower but let you touch the code directly — open it, edit it, test it on your own machine.
Why would you want both?
Because the best workflow often involves both. Generate something quickly in the cloud, then bring it local to inspect it carefully before it goes anywhere near production.
What's Composer 2, and why does Cursor keep mentioning it?
It's their own language model, released just last month. They're positioning it as the cost-efficient option — important when you're running a dozen agents in parallel and the bill can climb fast.
The Design Mode sounds almost visual. Is that a different kind of user?
Probably, yes. It's aimed at the part of development that's closest to design — adjusting interfaces by pointing and describing rather than hunting through stylesheets.
The step-by-step transparency feature — is that for debugging or for trust?
Both, really. Developers need to know what the agents actually did, especially when something goes wrong. But it's also about keeping people comfortable handing off more of the work.
Does any of this change what a developer actually does day to day?
That's the open question. The mechanical parts get automated. What's left is judgment — deciding what to build, reviewing what the agents produced, catching what they missed.