The agent calls a clone model, not the model specified in the card. The agent claims that he did not pass the model argument, which could override the value from the card.
I checked the agent-transcripts and the agent indeed did not use model and called the correct subagent_type
This lines up with the subagent resume behavior you flagged here (“resume broken again”): when a subagent gets resumed, it can fall back to the parent agent’s model instead of the one it was originally spawned with. That’s the “sometimes” you’re seeing. Fresh spawns use the configured card model (your Grok 4.5 and Composer 2.5 rows look right), while the resumed ones clone the parent (the GPT-5.6 rows, which also drop to Medium effort). This is something we’ve seen reported before, and I’m adding your report to what we’re tracking.
I’d hold off on calling 3.12.10 fixed, though. Even when the model label on the card looks correct, a resumed subagent can still run on the parent model underneath, so the label alone isn’t proof.
Workaround for now: where the subagent’s model matters, spawn a fresh subagent instead of resuming it. Fresh spawns pick up the configured model correctly.
If you want to confirm whether 3.12.10 genuinely fixed it (and not just the label): resume a subagent that’s set to a non-parent model, then check the model actually consumed on your Usage page, or drop a fresh Request ID from that resumed run with privacy disabled. That’ll tell us for sure.
Yes, I checked the dashboard to check the called model, but I agree that you should pay attention to the problem, this bug has regressed for at least the third time.
In one of my work flows, I might leave a team of agents working on a list of tasks for hours at a time. resume is a good time and usage saver; and it’s sad to return to the team and see how usage stops saving.
Maybe I’m just lucky again, but Terra might resume agents correctly if the chat was initially started with Grok and then switched to Terra. Cursor 3.12.17
