With the new support of UI-based testing in cloud/web agents, and the natural integration of background agents launched in Slack or GitHub, I would actually uninstall the Cursor IDE and do everything on the cloud.
There are many reasons for this:
It promotes async work. It’s simpler to shoot a prompt to each agent and loop through them. So is the background triggering through Slack, Github or event Cursor API…work ends up there to be reviewed in batches.
I can turn off my laptop and it will continue working. I can even lose it.
The UI is agent first, whereas the IDE is still code-first. So the natural flow is different, e.g. when reviewing units of work, the cloud enviroment promotes ignoring commits and reviewing everything later.
BUT the ONE thing that is blocking me from enjoying all of these is speed. The remote environment seems at least 10x slower. Boot up is one thing, dependent on the snapshot use, etc. But where it’s noticed the most is upon each prompt, I experience a 10s lag before it starts doing things.
Also, since you don’t see the diff progress as you do locally, and the UI is really bad at refreshing, you don’t really know if it’s just frozen or actually working. This makes the perceived lag even higher.
Please address the performance issues of the machines, the lag between prompts and the UI bugs where chats don’t open or freeze…
Hey, thanks for the detailed report. All three issues you’re describing are known:
Slow boot-up times, we’re tracking this.
Around a 10s lag between prompts, also on our radar.
UI freezing or not refreshing, known issue, especially when streaming environment logs. Your point about not being able to tell if it’s frozen vs. working is valid, the feedback loop definitely needs improvement.
No workaround to offer here unfortunately, but your report helps with prioritization. We’ll update here if there’s news.
Ok, tested it this week. I do see imrovements both in the first prompt/boot time and in the responsiveness or streaming of each response. This applies both when using Cloud through the website or through the new local Cursor Agents UI.
However, the locally ran model still works much faster and smoother. The gap is so high that it makes me default to Local, when I would default to Cloud otherwise instead because of the flexibility it provides.
I do applaude the progress made on transitioning from Cloud to Local and viceversa, this allows to get the best from both worlds, and is awesome. But one has to do it, and remember it and foresee next steps to avoid the bad moment of realising that you forgot to move the session back to the Cloud.