Hey, thanks for the report. In the end, the model decides how many sub-agents to run in parallel. The agent interprets your instructions and decides how many to launch. So explicitly adding “up to 4 in parallel” changes the behavior.
Right now the most reliable workaround is exactly what you found: specify the number in the instructions. I get that this isn’t ideal, since you’d want auto-scaling if the limit changes.
I’ve passed this to the team.
I also saw your related thread about batch scheduling Parallel subagents: next batch waits for entire previous batch to finish. It’s the same area to improve, and I’ll flag it too.
If you can share a Request ID from one of these sessions (chat context menu, top right, then Copy Request ID), that would help the team investigate the model behavior more precisely.