Grok strange quirks

Ideas Feedback
, ,
1

Hi, I’ve been using cursor for quite a long time now, and it was always getting better and better until… Grok… Is there something I’m doing wrong, something new about Grok that I’ve missed ?

Composer is good at almost everything, and even when it’s not good it’s not crap. I feel like Grok is good at nothing and is often plain crap.

Grok can spend 40 minutes on a problem without any good outcome that Composer or Opus can solve in just a few minutes (if not seconds!). Grok can’t stop editing everything that cross his path when it’s working on a subject, it’s like a junior engineer that can’t help but refactor every file it touches even if it works perfectly.

On top of that the number of non-building results I have with Grok shocks me. That is something that almost never happens with Composer. Globally the code quality is not good, here is an example from this morning :

t('aiAgent.streamingPlaceholder', {
  defaultValue: locale === 'en' ? 'Reviewing your file…' : 'Consultation de votre dossier…',
})

We have a normal i18n implementation, Composer always create the keys in the right place and uses them correctly.

Aside the pure code quality, Grok seems to have a lots of quirks when working with git. It sometimes create new branches and merge requests for each new interaction we have in a session, this usually happens when we interract with cloud agents through linear, slack, or directly in the sessions through cursor’s UI.

Don’t get me wrong, I love cursor, but for now I’m not a big fan of Grok in Cursor.

After this little bit of rantig, my question is there something I’ve missed, does grok have a specific way of working that I’m not aware of ?

5

Hey, thanks for the detailed report. There are actually two separate things here, so I’ll go through them one by one.

On code quality. There’s nothing special you need to do to “prepare” Grok 4.6 differently than Composer. It’s just a different model with a different style. Your i18n example with defaultValue inline is exactly the kind of thing that’s useful for the team. The most effective steps are:

  • Give a thumbs down on runs where the model went off track. That flags the specific run for review.
  • Share your notes in the Grok 4.6 thread: Share your Thoughts on Grok 4.6
  • Add a project rule for your i18n convention, like where keys live and no inline defaultValue. That helps any model.
  • Check the effort level and the Fast toggle in the model picker. Cloud Agents often default to the Fast variant, and that can explain “40 minutes with no progress”.

On git. What you’re seeing in Linear isn’t Grok. That’s how Linear agent sessions work, and it happens with any model.

Each new top level comment that mentions Cursor, or each time you delegate an issue, creates a new agent session. New session means a new Cloud Agent, and it creates its own branch and its own merge request. To continue the same session:

  • In Linear, open the issue and find the thread where Cursor posted updates, like the one that says “Started cloud agent …”.
  • Reply inside that thread, not as a new comment at the bottom of the issue.
  • Or open the agent at cursor.com/agents or in the Agents view in Cursor, and send your follow up from there. Those follow ups continue work on the existing branch and update the current MR.

Same in Slack. Reply in the agent thread, not as a new top level message.

If you reply inside the existing thread and it still creates a new branch and MR, send the agent link like cursor.com/agents/… and I’ll take a look at that specific run.