There’s a particular awkwardness when a well-loved product stops being what it was, but hasn’t yet become what it wants to be. Cursor 3.0—a totally rebuilt version of Cursor centered around agent-orchestration—out today, is in that weird in-between stage: There are many promising early signs, but it’s not yet mature enough to merit a switch.
Cursor helped create the AI-native integrated development environment (IDE). It was the product that made “coding with AI in your editor” feel like a genuine category. But the thing that made them a category leader has turned out to be constraining.
The rise of Claude Code shifted the dominant coding paradigm away from pair programming—which requires a code editor—toward agent delegation and orchestration—which doesn’t. That left Cursor in a tough spot.
Do they satisfy their core IDE audience and continue building on top of an editor? Or do they start from scratch and build an agent-orchestration tool like Claude Code, Codex, and Conductor?
With Cursor 3.0, they’ve decided to take the plunge toward the latter: This release is a rebuilt product made for dispatching, monitoring, and managing AI agents rather than manually writing code. We’ve been testing it internally for about a week, and there are a lot of promising signs. It’s extremely fast and resource-light, it has an impressive local-to-cloud programming system, and its built-in model, Composer 2, is quick, smart, and concise.
But as a product, it’s still too early in its lifecycle to be a serious contender against the more mature agent orchestration offerings from Anthropic and OpenAI.
The question that kept coming up during our testing was, “Who is this for?” Power users who already work in Claude Code or Codex won’t use a new agent orchestration layer unless it’s 10 times better than what exists, because they already have one that works well. And existing Cursor users who love the IDE are likely to be upset that it has been deprioritized.
That’s why we believe Cursor is in an awkward in-between stage. They’ve made the right strategic move to prioritize agent orchestration. But it will take some time before what they have is truly competitive. Their team seems to be iterating incredibly fast, so we’ll be paying attention over the coming weeks and months as it improves.
‘Who is this for?’
Won’t use a new agent orchestration layer unless it’s 10x better than what exists.
Likely to be upset that the IDE has been deprioritized.
Thanks to our Sponsor: MongoDB
AI data looks nothing like a spreadsheet
Some of our legacy tools are holding us back. AI applications generate unstructured data, like text video, audio, sensor feeds, and user interactions. But traditional databases were designed for structured rows and columns. That’s where MongoDB Atlas comes in. It natively handles JSON-like unstructured data with the flexibility to store, query, and scale diverse datasets without rigid schemas. Plus native vector search, auto-scaling for demand spikes, and multi-cloud deployment across 125+ regions make it the perfect fit for the AI era.
What’s new
Cursor 3.0 features a new default interface built from scratch—a departure from the VS Code fork that the product was originally built on. The company says it keeps the most useful pieces from earlier versions (the browser, diffs, file, and plan viewing) while adding support for patterns that reflect how coding with AI has changed. Highlights according to the Cursor team:
Before
Editor with AI
Now
AI with an editor
- Cloud and local agents that can hand tasks back and forth
- Cloud agents that run on their own machines
- Working across multiple workspaces
- Agent-first git management, or keeping track of changes to a project’s files
- A new model picker that lets users set defaults based on their preferences, with access to any frontier model, including Cursor’s own Composer 2
The Cursor team notes that the editor is still present, just no longer center stage. In practice, the shift is bigger than that framing suggests. The default view is an orchestration panel—repos on the left, a chat-driven workflow in the center, browser and markdown editors docked alongside. The file explorer is gone from the default layout. Where Cursor used to be an editor with AI built in, it’s now an AI dashboard with an editor tucked behind it.
The previous iteration, Cursor 2, featured a three-pane interface with a file tree on the left, an editor in the center, and an AI chat panel on the right. (Image courtesy of Katie Parrott.)
Cursor 3.0’s default view opens on an agent-centered workspace, with repos in the sidebar and a chat-driven orchestration panel replacing the traditional editor-first layout. (Image courtesy of Dan Shipper.)
Cursor charges a base price per seat. But on top of that, some AI usage is metered—the more you use it, and the more expensive the model you pick, the more you may end up paying. Kieran Klaassen said he spent roughly $2,000 in the course of testing (comped by the Cursor team). For heavy users, the cost may end up looking very different from the sticker price.
The Reach Test
“Things I love: Cursor clearly knows how to build a desktop app. Cursor 3.0 is fast and light as compared to resource-intensive monsters like Claude Code Desktop and Codex. It also has a promising cloud orchestration layer: You can spin up an agent from your desktop, walk to get coffee, and watch a video demo from a cloud walking you through the feature it just built. But, ultimately, it’s just not mature enough for me to switch. This is the right strategic direction for Cursor, but it will take a few more releases (and some pain) before this product becomes a viable alternative.”
“Cursor invented the AI-native IDE, and it deserves credit for that. 3.0 feels unfinished—the harness is too aggressive, sessions break, and the interaction model is inconsistent. But they’re moving fast, and if they smooth out the rough edges, this could be a genuine alternative to CLIs for people who aren’t comfortable in a terminal. One thing to flag: I spent roughly $2,000 in two days of normal work—the kind of usage I get for a flat monthly fee with Claude.”
“This moves away from the thing I still valued Cursor for: an IDE where I could micromanage the things AI failed at. The new orchestration layer doesn’t offer anything I can’t already get from my Claude Max plan, and the pricing math makes it worse—$200 per month for unlimited Opus versus paying per-token in Cursor for a similar experience. They are a smart team, and I’m confident this is the direction their customers are pulling them in, but it’s not aimed at me.”
“The app is faster, and the rebuild is noticeable. But I’m not reaching for it—I still open Cursor as an IDE because I’m used to it, not because 3.0 offers something new. Missing basics like a branch selector doesn’t help.”
The headline findings
Subscribers only
Become a paid subscriber to Every to unlock this Vibe Check and learn about:
- What works: The one feature every tester agreed was good
- What doesn’t: The small UI decision that frustrated a power user
- Bottom line: Why the thing that makes Cursor 3.0 ambitious is what makes it vulnerable
The identity gap shows up in missing basics
Arc lint, node haze, frame tethered; looped soft and landed wide. The line held, the hinge set, the lock slid early. Trace spool, log fog, seam drop; path locked, run shipped.
Multi lanes sparked, then braided back without prompts. The flow stayed intact, the blocks aligned, the bridge held.
The harness fights the model
Trace spool, log fog, seam drop; path locked, run shipped, no extra pulls. The mesh held, the marks lined up. Arc lint, node haze, frame tethered.
Long arcs, cadence banded, next move early. The trace held its line under load.
Desktop app performance
Frames the task, trims edge noise, keeps the lane clean end to end.
Composer 2 earns its spot
Spins multiple passes, braids the best line, keeps output stable.
Local-to-cloud orchestration
Finds adjacent tools, shifts tone, lands clean without perfect prompts.
The identity gap shows up in missing basics
The arc slows, the loop elongates, and the cadence slips between passes.
The harness fights the model
Contours drift in visual builds; polish lands uneven across frames.
Session persistence is fragile
Occasional claims outrun the ground truth beneath them.
The bottom line
Arc and lane converge; the trace holds under pressure, the seam lands clean.
Get all of our AI ideas, apps, and training
Every is the only subscription you need to stay at the edge of AI—trusted by 100,000 builders.
Expert led courses and camps
Four productivity apps
A Discord community learning together