On April 2, Anysphere shipped Cursor 3 - a ground-up rebuild of the AI coding platform that millions of developers rely on daily. This isn't a new theme or a faster autocomplete. The team built an entirely new interface called the Agents Window, designed around a thesis that would make most developers uncomfortable: your job isn't to write code anymore. It's to manage the things that write code for you.
From Copilot to Control Tower
The shift here is structural, not cosmetic. Previous versions of Cursor were essentially VS Code with a very good AI assistant bolted on. You'd tab-complete, chat with the model, maybe let it refactor a function. Cursor 3 keeps all of that, but adds a parallel workspace where developers can spin up multiple autonomous agents, each working on different tasks across different repositories simultaneously.
Think of it less like pair programming and more like running a small engineering team. You assign tickets. Agents work in isolated git worktrees. They produce diffs, screenshots, and demos for your review. You approve, request changes, or redirect - exactly like a tech lead reviewing pull requests from junior engineers.
The numbers back up the bet. According to Anysphere, agent usage grew over 15x in the past year. In March 2025, autocomplete (Tab) users outnumbered agent users 2.5-to-1. That ratio has completely flipped - it's now 2-to-1 in favor of agents. Internally, 35 percent of pull requests merged at Cursor itself are created by autonomous cloud agents.
What's Actually New
The headline features tell the story of where Anysphere thinks development is headed. The Agents Window is a standalone interface for orchestrating local and cloud agents from a unified sidebar. Agents can be kicked off from desktop, mobile, Slack, GitHub, or Linear - and handed off between local and cloud seamlessly.
Design Mode lets developers visually annotate browser elements and point agents to exact UI components instead of describing them in text. The /best-of-n command runs the same task across multiple models in parallel worktrees, then lets you compare results side by side. Bugbot, their automated PR reviewer, now goes beyond flagging issues - it spins up a cloud agent, tests a fix, and proposes it directly on the pull request. According to Cursor, over 70 percent of Bugbot flags get resolved before merge, and 35 percent of its automated fix suggestions are being merged into production code.
For enterprise teams, there's a Helm chart and Kubernetes operator for scaling to thousands of agent workers, with self-hosted options that keep code and execution entirely within your network.
The Developer Backlash Is Real
Not everyone is celebrating. The Hacker News discussion around the launch surfaced a sharp divide. Several developers described the agent-first approach as premature - one called it being forced into managing agents rather than writing code. Others reported spending upwards of $1,800 per month on Cursor credits for heavy agent usage, then switching to Claude Code or Codex at a fraction of the cost.
The pricing concern is legitimate. Cursor Pro starts at $20 per month, but premium model usage draws from a credit pool that power users can burn through quickly. The Ultra tier runs $200 per month. Meanwhile, according to Menlo Ventures, Claude Code has captured roughly 54 percent of the AI coding market with its terminal-first approach and simpler pricing.
There's also a trust question. Cursor's proprietary Composer 2 model was revealed to be based on Moonshot AI's open-source Kimi 2.5 without upfront disclosure - a stumble that eroded goodwill with the developer community that values transparency.
Why This Matters Beyond Cursor
Whether you love or hate the direction, Cursor 3 is a leading indicator of where the entire industry is headed. GitHub Copilot is adding agent capabilities. Anthropic's Claude Code already supports autonomous coding sessions. OpenAI's Codex continues to climb benchmarks. The question isn't whether developers will manage AI agents - it's how soon and how much control they'll retain.
The companies that went from zero to $1 billion in annual recurring revenue in three years tend to be onto something. Anysphere hit that milestone in December 2025, backed by $3 billion in funding from Thrive Capital, Nvidia, and Google at a $29.3 billion valuation. They're not guessing about the future of development. They're forcing the bet.
What To Do About It
1. Try the Agents Window without abandoning your workflow. Cursor 3 keeps the traditional IDE intact. Open the Agents Window alongside your editor and delegate one low-stakes task - test generation, documentation, or a refactor - while you keep coding normally.
2. Benchmark your actual costs. If you're on Cursor Pro or Ultra, track your monthly credit consumption for two weeks. Compare against Claude Code Max ($100/month unlimited) or Codex. The right tool is the one that delivers quality output at a price your team can sustain.
3. Start with Bugbot on your GitHub repos. Automated PR review with a 70 percent resolution rate is free productivity. Even if you skip the rest of Cursor 3, this alone can catch bugs your team misses under deadline pressure.
4. Watch the agent-to-Tab ratio at your org. If your team's AI usage is still mostly autocomplete, you're one generation behind. Start experimenting with autonomous coding sessions - in any tool - to build the muscle memory for where this is going.
HRIM's Take
We build software for a living, and we're watching this shift closely. Cursor 3's bet is directionally right - agents will write the majority of production code within the next two years. But the execution has rough edges: opaque pricing, a proprietary model built on undisclosed open-source foundations, and a vision that's running ahead of what most development teams are ready for. The smartest move right now isn't to go all-in on any single tool. It's to get comfortable managing AI agents as part of your workflow, regardless of which platform you use. The developers who learn to orchestrate effectively will build 10x faster. The ones who resist will find themselves managing the agents anyway - just later, and less prepared.