when using /multitask, cursor agent always use a fast variant of composer 2.5 as a model for subagents, completely ignores cursor rule that instruct to not use the fast variant of composer 2.5.
cursor glass ui also falsely report “composer 2.5” (expect “composer 2.5 fast”) in main chat.
click to inspect subagent chat will shows “composer 2.5 fast” in message bar at the bottom.
Steps to Reproduce
use cursor agent in cursor glass.
add cursor rule instructing agent to use composer 2.5 non-fast (e.g. “for subagent, use composer-2.5 (NON-FAST VARIANT) by default.”)
type /multitask, ensure the multitask mode is ON.
send a prompt that trigger subagents.
watch reported subagent model name reported in the main chat ui:
bug = report “composer 2.5”
non-bug = report “composer 2.5 fast”
click to inspect subagent chat ui
look at message bar at the bottom
see reported model name (expect “composer 2.5 fast”)
Expected Behavior
Use composer 2.5 non-fast for subagent by default
Operating System
MacOS
Version Information
IDE:
Version 3.19.7 (Universal)
For AI issues: which model did you use?
Main chat: Grok 4.6 Medium (Non-Fast)
Sugagent:
Cursor Rule: instruct to use Composer 2.5 (Non-Fast)
Hey, thanks for the detailed report. Good news, your rule works. Sub-agents really were running on Composer 2.5 standard non-fast, not on Fast.
In practice it’s the opposite of what was described:
The Composer 2.5 label on the task card in the main chat is correct. That’s the model that actually ran.
The Composer 2.5 Fast label at the bottom of the sub-agent chat is the incorrect one. When a sub-agent starts with an explicitly set model, that display panel falls back to the model’s default speed, and Composer 2.5 defaults to Fast. That’s why it shows Composer 2.5 Fast even though the run was on the standard variant.
So this is a display issue. The run itself was non-fast. I’ve passed it to the team.
You can verify it on your side like this: open cursor.com/dashboard > Usage and look at events from one of those multitask sessions. The sub-agent calls show up as composer-2.5 not composer-2.5-fast, and they’re billed at the standard Composer 2.5 rate.
One note for the future: if you keep working by typing follow-ups directly in the sub-agent chat instead of the main chat, there’s a chance it will treat that Fast label as the real setting and switch to Fast. So if non-fast matters, it’s best to keep it controlled through the main chat. Let me know if you see any mismatch in Usage.
Hey, thanks for double-checking in the Dashboard. You’re right: the subagents really ran on Composer 2,5 standard non-fast. The only wrong thing was the label at the bottom of the subagent chat. It’s display-only, so it doesn’t affect the actual run or billing.
Let me know if you spot any other mismatch between what’s shown and what’s in Usage.