Cursor 3 Released: AI Coding Tools Can Finally "Work Independently"

Honestly, when I saw the news about Cursor 3’s release, my first reaction was — here comes another “pie in the sky.”

But after watching the demo video, I was stunned.

This isn’t simple code completion, nor is it Claude Code’s “you say one sentence, it does one step” rhythm. Cursor 3’s “Glass” feature is true multi-Agent collaboration — you can open 3 Agents simultaneously: one writing code, one writing tests, one running CI, and they communicate with each other and fix bugs themselves.

It’s like going from “one person coding” to “directing a 3-person team.”

My personal feeling is that this transformation is more fundamental than pure speed improvements. It changes the organizational structure of how development happens.

But seriously, is this “multi-Agent collaboration” really reliable? I’ve seen many AI coding tool demos, but they always fail in actual use — context understanding goes off track, tool calls fail, generated code looks good but crashes when run. Can Cursor 3 be different this time?

What interests me is that Cursor 3 made a smart design choice: seamless switching between cloud and local Agents.

This means you can hand computationally intensive tasks (like analyzing an entire codebase, generating large refactoring plans) to cloud Agents, while keeping latency-sensitive tasks (like modifying a function, completing a few lines of code) local. This “division of labor” logic is more pragmatic than “all cloud” or “all local” approaches.

Honestly, I’m kind of curious to try this “multi-Agent collaboration” feeling. Imagine sitting there: on the left screen Agent A is writing business logic, on the right screen Agent B is writing unit tests, in the middle Agent C is running integration tests — you’re like a project manager, occasionally chiming in “change this variable name” or “think about that boundary condition again.”

It sounds great, but on second thought, it’s a bit strange.

If AI can write, test, and fix by itself, what’s the developer’s role? From “person who writes code” to “person who manages Agents”? Or are we just handing off the “grunt work” to AI while we handle the “brain work”?

My personal judgment is — at least in the short term, this multi-Agent collaboration is more like an “amplifier” rather than a “replacer.”

It can automate repetitive tasks like “writing tests” and “running CI” for you, letting you focus on architecture design, business logic, and technology choices where judgment is needed. But if you expect it to think for you about “should this feature be built” or “is this architecture reasonable,” you’re expecting too much.

That said, Cursor 3’s multi-Agent collaboration does show me a possibility — AI coding tools are evolving from “assistants” to “teammates.”

Previous Copilot and Cursor 2 were more like “you ask, it answers” assistants; Cursor 3’s “Glass” starts to have the embryonic form of “working independently.” It’s like going from “co-pilot” to “partner.”

Of course, this might just be my over-interpretation of the demo. Real results require hands-on experience.

Final question: If AI can really open 3 windows to write code, test, and fix bugs by itself, do you think this is “liberating your hands” or “a prelude to being replaced”? Would you let it run the entire process, or would you watch every step?

Anyway, I’m already in line for Cursor 3’s beta access. Once I get it, I’ll share my real experience — including the failures.