Kelden Lin ᯅ

@aye_aye_kaplan i haven’t experienced the same :( it struggles with longer horizon tasks, and it gives me the “it’s a good subagent” vibe. i really wish it could do longer horizon tasks! feels like it stops short of being thorough and perhaps that’s a result of optimizing it as a subagent

It’s an excellent model. My personal opinion is it’s near Opus quality (indistinguishable to me for most tasks), but way faster and way cheaper. Enjoy!

@haydenbleasel banger

if you want to make sure applicants aren't using cluely just ask them to close their eyes while answering technical questions

@leerob i might be crazy but did you guys remove the cmd+shift+space shortcut for voice input? (for agent window) is there a particular reason for the change? cmd+k+"voice" just doesn't hit the same :(

Meet the new Cursor! Very excited about this. Wanted to share a bit more of the story of how we landed here, how the product evolved, and some of the technical details on the new interface. I've been coding primarily with agents since Opus 4.5, but hadn't found an interface I loved (including our own). Agent sidebars or CLIs worked but still felt limiting to me. And our first iteration of the "agent window" wasn't good enough yet. So we went back to the drawing board to build a completely new interface for agents in December. The trend seemed pretty clear that increasingly less time would be spent in traditional IDEs. But as we started to dogfood early versions, it was very hard to give up some parts of an editor. Even if agents write 98% of the code, that last 2% of viewing files, debugging, many small edits and refactors, and having all the niceties like go to definition, LSPs, and more were really important. We couldn't remove those. So @ryolu_ and I started prototyping some ideas late Dec for a new interface. It would start simple/zen, but then allow you to still go deeper as needed. And slowly we developed enough conviction to make it real. The Cursor eng/product team then took some of those early ideas and made something 10x better than I imagined. Seriously major kudos to the team! We started fresh with this new UI in a lot of ways. "Deleting the product" is especially important as models continue to improve and the UX needs to be continually rethought. However that doesn't mean you have to throw out *all* the good ideas. Making it easy for existing users to adapt is also very important. In this new interface, we own all the pixels. We were able to design a system/architecture that takes all the learnings from Cursor 1/2 and moves away from some of the VS Code constraints we were limited by. I definitely empathize with feedback that in Cursor 2 we were moving around the UI too much and changing icons/buttons. Agents were taking over more and more work, and they started to break out of the IDE UI. We needed to iterate and try a bunch of things, and that was annoying for those of you expecting a more consistent editor experience. Making this new agent interface as a separate window actually also makes the Cursor 2.0 IDE *better*! Rather than continuing to try and extend the IDE to have agents own the entire UI, we were able to simplify and delete a lot of code by using existing VS Code patterns. Namely, agent chats are now just normal tabs like any other file. This is much more stable and familiar for doing splits/panes and all related keyboard shortcuts. A win-win-win, as they say. But also in this architecture refactor, we were able to address some local vs. cloud divergence and tech debt that had accumulated over time. The core Cursor agent harness is the same across the desktop app, web app, CLI, etc. So there really shouldn't be two code paths like: if (local) { ... } else if (cloud) { ... } Cloud agents were not used much until we gave them the ability to use a computer and record demos of their work, so now that usage has grown considerably in the past few months, it was even more important to nail this abstraction. We think cloud usage will continue to grow and be a big part of 2026. Finally (this is already a long post, oops), we have been able to really focus on performance in the new interface. I'm sure there will still be things to improve (please send them to us) but we've spent considerably more time profiling, investigating, and patching memory/cpu leaks. We are also now using the React Compiler! s/o @potetotes who has also been making a bunch of perf improvements. The end result here is that Cursor 3 feels much more pleasant to use. Faster, more reliable, less UI jank. You get to use all your favorite models, local or cloud, run automations, install plugins, get back demo videos, and more. Give it a try and lmk your feedback! We're gonna be shipping updates quickly in the coming days.

@haydenbleasel @OpenAI hell yeah!!

I've joined @OpenAI as a Member of Technical Staff. AI will shape the future of humanity in profound ways. I believe those of us who can help build it have a responsibility to help steer it in a direction that benefits everyone, for all humankind. The future is bright, but it takes hard work and care to make it that way. Very excited to play my part.

guys i might've vibe coded too hard..

@haydenbleasel yo thank you for all the amazing libraries it’s really insane what you’ve done!!

Today is my last day at Vercel. I've had an absolute blast over the last year. I got to work on a lot of my own projects that helped form the basis for our AI Cloud, helped other teams ship their great ideas, and worked alongside some of the most talented people I've ever met. I'll have news on my next adventure for you on Tuesday but for now, wanted to reflect on my time here and the projects I worked on. Read on if you're interested ↓

@ryolu_ @robinebers is it composer 2 time?

@robinebers a new world awaits check back today to see what we’re cooking

sometimes I think about leaving Cursor behind 🫠 it's still the best tool but I had so many conversations lately with people that feel the same yes, it's the best AI agent there is but the direction feels unclear, I sense a lack of vision wish I'd know what their grand plan is, or if they have one coming from an engineering and product owner background, this hits harder for me than others maybe, but I really hope they figure it out

@haydenbleasel can’t wait to try out stagger

▲ Streamdown 2.5 just dropped. This release is all about streaming fidelity - smarter math rendering, better animations, and a bunch of fixes to make your AI chat UX feel rock solid. Details below ↓

@steipete @OpenAI @openclaw absolute legend

I'm joining @OpenAI to bring agents to everyone. @OpenClaw is becoming a foundation: open, independent, and just getting started.🦞 steipete.me/posts/2026/ope…

@BLKMDL3 any updates for hw3 or this just v14?

Huge wave of 2025.45.9.1 rolling out. Check your apps!

@OpenAI holy

GPT-5.3-Codex-Spark is now in research preview. You can just build things—faster.

@DirtyTesLa hw3 here, was surprised at how good it was.. the only annoying thing is that it doesn’t know don’t turn right on red signs and it doesn’t move over for highway exists early enough sometimes (for my comfort). how are those in fsd 14?

What is FSD still doing that is annoying you? Or is every drive perfect while you relax in the driver seat "supervising"?

@hellokillian let’s goooooooooo every post i see about clawdbot and claude code i think of open interpreter and how yall did it years ago

❤️ Alright alright I’ll drop something next week.

@tejashaveridev @cursor_ai wow that’s super cool! thank you for the detailed write up. unfortunately can’t test it on my end cuz im assuming that’s on nightly or early testing? i’m in our companies enterprise plan so we can’t get anything outside of stable updates 😞

Long running agents! and in terms of task, i did have a lot of ambitious massive refactor type of tasks in backlog for a while, and so far whatever i throw at it its been really good with fixing, then testing for regressions and code quality is pretty good as well tasks ive done so far thats taken over a 6 hours each: -refactor rbac system entirely -enable stricter typecheks -gut and remove a LOT of legacy code -optimize and break down big files/code its not just that it can go for a while, its i can give a task and actually come back and see its really taking its time and going through literally every line of code and making sure nothing gets left out - that was a big problem i was seeing, the bigger task you go for the more likely it wont do parts of it or will be lazy and skip it "for a later date" interesting part is how this harness is doing context management, hope they do an article on this soon

7 hours and still counting, man the fact that this is now possible is just mind blowing @cursor_ai upcoming feature: long running agents

@cursor_ai love the consistent updates and share outs on the long running agents, these are super informative!

We've been working on very long-running coding agents. In a recent week-long run, our system peaked at over 1,000 commits per hour across hundreds of agents. We're sharing our findings and an early research preview inside Cursor.

am i crazy but how can we use multiple skills in codex in a single prompt? / hint goes away after my first skill and i just have to kinda verbally prompt it? i want a consistent way to invoke it

@ericzakariasson @gregorian909 @cursor_ai can we also add an option to build in worktree? i sometimes find myself planning on main and building in worktree. also if im at a worktree and build with new agent does it do it in a worktree or main?

@gregorian909 @cursor_ai try this!

It'd be really helpful to create a plan locally and then hand it off to the cursor cloud agent @ericzakariasson @cursor_ai

@haydenbleasel what did you use to make this video?

One command for better AI interfaces, with AI Elements Skills. ▲ ~/ 𝚗𝚙𝚡 𝚜𝚔𝚒𝚕𝚕𝚜 𝚊𝚍𝚍 𝚟𝚎𝚛𝚌𝚎𝚕/𝚊𝚒-𝚎𝚕𝚎𝚖𝚎𝚗𝚝𝚜

@haydenbleasel how r u always cooking

AI Elements has its own website now!