Cursor 3 bets $29B that developers want to manage agents, not write code

Anysphere just ripped out the heart of Cursor and replaced it with something completely different.

On April 2, Cursor 3 launched — and it’s not an IDE update. The company dropped its code-editor identity entirely and rebuilt the product as an agent orchestration platform. You no longer open Cursor to write code. You open it to dispatch, monitor, and manage a fleet of AI agents that write code for you. Desktop, mobile, web, Slack, GitHub — pick your surface, spin up an agent, and walk away.

This is a $29.3 billion company telling its users: stop typing, start delegating. With 517 points and 384 comments on Hacker News the same day, plus 226 upvotes on Product Hunt, the launch made noise. TechCrunch and Gizmodo both covered it. But the real question isn’t whether people noticed — it’s whether Cursor can pull off this pivot fast enough to stop the bleeding to Claude Code.

From editor to command center

The centerpiece of Cursor 3 is the Agents Window — a full-screen workspace that replaces the old Composer pane. Think of it as mission control for AI coding agents. You can run multiple agents simultaneously, each in its own panel: one refactoring a service layer, another writing unit tests, a third handling a cloud deployment. Each agent operates in an isolated Git worktree, so they can’t overwrite each other’s work. The worktrees share the same Git object store, keeping disk usage manageable even when you’re running half a dozen agents in parallel.

The agent infrastructure spans local machines, cloud VMs, and — as of March 31 — your own private infrastructure. Self-hosted cloud agents let enterprise teams keep their codebase, build outputs, and secrets entirely within their own network. Workers connect outbound via HTTPS to Cursor’s cloud — no inbound ports, no VPN tunnels, no firewall changes. Up to 10 workers per user, 50 per team. For Fortune 500 companies handling sensitive code (and Cursor claims more than half the Fortune 500 as customers), this is table stakes.

Then there’s Design Mode. Instead of describing UI changes in text, you click and annotate elements directly in the browser. Point at a button, tell the agent what to change, and it executes. It’s the kind of feature that makes you wonder why nobody built it sooner — visual feedback for visual problems, no translation layer required.

The /best-of-n command is the quietly powerful addition. Run the same prompt across multiple models simultaneously, each in its own worktree, compare the results side by side, and pick the winner. A refactoring task might produce cleaner code from Claude, but better test coverage from GPT-5.4. With best-of-N, you don’t have to guess. You just run all of them. This is the first major coding tool to make model selection a runtime decision rather than a settings choice.

The Kimi shadow still looms

Cursor 3 doesn’t launch in a vacuum. It arrives two weeks after one of the most damaging trust incidents in AI tooling history.

On March 22, Cursor admitted that Composer 2 was built on top of Moonshot AI’s Kimi K2.5 — a Chinese open-source model — without disclosing it in the original announcement. An X user named Fynn reverse-engineered the model and found Kimi’s fingerprints in the code. Cursor’s VP of developer education Lee Robinson acknowledged “Yep, Composer 2 started from an open-source base!” and co-founder Aman Sanger admitted it was “a miss to not mention the Kimi base in our blog from the start.”

This matters because Cursor’s user base is technical, and technical users care about model provenance for practical reasons. Data routing, privacy, jurisdiction — if your coding queries are flowing through a model with Chinese origins, enterprise compliance teams want to know about it. Making it worse, this wasn’t the first time: Composer 1 used DeepSeek’s tokenizer without disclosure either. Once is an oversight. Twice is a pattern.

Cursor 3 needs to rebuild that trust. The self-hosted cloud agents, the transparent model selection through best-of-N, the explicit multi-model architecture — these aren’t just features. They’re answers to the question: “Can I trust you with my code?” Whether the answer is convincing enough depends on whether Anysphere’s actions match its new transparency messaging going forward.

The Claude Code problem

Here’s the competitive reality Cursor 3 is responding to: Claude Code has been pulling developers away from Cursor at a pace that should worry any investor with skin in that $29.3 billion valuation.

By January 2026, Claude Code captured 31.6% market share, overtaking GitHub Copilot in just eight months. Developer satisfaction tells an even starker story — 46% of developers name Claude Code as their “most loved” tool, more than double Cursor’s 19%. Claude Code hit an estimated $2.5 billion run-rate by early 2026. These aren’t experimental users kicking tires. These are developers voting with their wallets.

Cursor’s counter-argument is revenue: $2 billion in annualized recurring revenue as of March 2026, doubling in just three months, with 60% from enterprise contracts. That’s not a company in crisis. It’s a company in transition. The AI coding market has fragmented into distinct lanes — GitHub Copilot holds 37-42% enterprise market share by headcount, Claude Code dominates developer satisfaction, and Cursor leads in commercial traction. Nobody has a lock on all three.

Cursor 3’s bet is that the future of coding tools isn’t about which model is best — it’s about which platform orchestrates models best. By going model-agnostic with best-of-N and letting agents run across Claude, GPT-5.4, Gemini, and its own Composer 2, Cursor is positioning itself as the layer above the models. If Claude Code is the best hammer, Cursor wants to be the construction foreman who knows when to use which hammer.

$2B ARR and the pressure behind the pivot

Let’s talk about the elephant: Anysphere raised $2.3 billion in its Series D from Nvidia, Google, and others, pushing the valuation to $29.3 billion. That’s roughly 15x its current ARR — aggressive even by AI company standards.

The revenue growth is real. Going from $1 billion to $2 billion ARR in a single quarter is the kind of trajectory that makes Series D investors sleep well. But sustaining it requires Cursor to be more than a coding assistant. It requires Cursor to be the platform where all AI-powered development happens — the workspace, the orchestration layer, the interface between human intent and machine execution.

That’s exactly what Cursor 3 is designed to be. The Agents Window isn’t a feature. It’s a platform thesis. If developers adopt the agent-orchestration workflow, Cursor becomes sticky in a way that a code editor never could be. You don’t switch away from your command center the way you switch away from an editor. The multi-repo support, the cloud agent infrastructure, the Slack and GitHub integration points — they all serve the same goal: make Cursor the single pane of glass for AI-powered development.

The risk is equally clear. The Hacker News thread surfaced it immediately. One top comment read: “I wish they’d keep the old philosophy of letting the developer drive and the agent assist.” A Cursor engineer clarified that the traditional IDE still exists alongside the Agents Window — you don’t have to give up manual coding. But the product messaging is unmistakable. Cursor is telling developers where it thinks the future lives.

Two camps, one market

Cursor 3 is the clearest signal yet that the AI coding tool market is splitting into two camps. On one side: tools that make individual developers faster — Claude Code, GitHub Copilot. On the other: platforms that turn developers into agent managers — Cursor 3, and presumably whatever GitHub builds next.

The pricing structure tells you which camp Cursor is betting on. Pro at $20/month, Pro+ at $60, Ultra at $200. The old “fast requests” model is gone, replaced by token-based billing where cost depends on which model you use and how complex the task is. If you’re running four agents across two repos using best-of-N with three models each, you’re burning through tokens fast. The Ultra tier starts to make a lot of sense — and that’s exactly where Cursor wants its power users to land.

The irony is that Cursor got to $2 billion ARR as a coding assistant — the thing it’s now telling users to move beyond. Two weeks after the Kimi trust crisis, with Claude Code eating into developer mindshare and a $29.3 billion valuation to justify, Cursor 3 is Anysphere’s biggest product bet. Not an incremental update. Not a new model. A complete reimagining of what a coding tool should be. The agents are running. The question is whether developers are ready to stop writing code and start managing the things that do.

Top AI Product

Leave a comment Cancel reply