Hi,
I think there’s a bug with Explore Subagent Model.
With Grok 4.7 Extra High:
First subagent: Grok 4.7 Extra High
Nested subagent: Auto
With Grok 4.6 Extra High:
First subagent: Grok 4.6 Extra High
Nested subagent: Grok 4.6 High Fast
Grok 4.6 still changes the reasoning mode, which is a bit odd, but at least it keeps the same model family.
With Grok 4.7, it falls back completely to Auto.
I’m attaching both screenshots for comparison.
Thanks!
Steps to Reproduce
a subagent launched by another subagent
Expected Behavior
All subagents, including subagents launched by other subagents, should use the configured Explore Subagent Model instead of falling back to Auto.
Hi @Rodolfo_Hoffmann , thanks for the clear report and the screenshots. This is a known issue we’re tracking. When your main chat is on Grok 4.7, subagents get switched to Auto instead of staying on Grok 4.7. Nothing in your setup is causing it.
From the request you shared, it actually starts with your first subagent. It also ran on Auto, even though its label showed Grok 4.7 Extra High. The nested subagent then picked up Auto from it, which is why Auto only shows up one level down.
For now, running your main chat on Grok 4.6 Extra High keeps every subagent level on Grok 4.6 Extra High. I checked your Grok 4.6 runs, and that’s what happened, even though the nested row was labeled “High Fast”.
One small note: the Explore Subagent Model setting only affects the built-in Explore subagent. Your coordinator and developer subagents use the model of your main chat, and that handoff is where Grok 4.7 gets switched to Auto. We’ll let you know when we have an update here!
Quick update: the issue where subagents launched from a Grok 4.7 chat were switched to Auto has been fixed on our side, so no update should be needed.
You can switch your main chat back to Grok 4.7 Extra High, and subagents (including nested ones) should now stay on Grok 4.7. Let me know if you still see any of them land on Auto.
the issue is kinda clear , conflicts in the models internals slugs. I was able to avoid that when i forced the actual models slug.
it is obvious that the model gets confused between cursor-grok and grok then it falls back to auto. fix the issue by dismantling the cursor setup. It is not transparent.
examine the cursor SQL db itself !
The actual model slugs usually matches the one posted on the web usage page