gpt-5.3/4 often refuse to use subagents

[buz-ch]

For some reason, gpt-5.3 and 5.4 still end up doing a lot of stuff in the main session instead of branching into subagents. It some times does, but definitely less than it should, e.g., I've caught it planning in main session instead of invoking plan-phase in a subagent multiple times.

Since the model is smart and has a huge context window, it usually works out ok but it's far from the intended approach.

Other models seem more open to use subagents but when prompting manually, I've found glm to use them a lot more when explicitly told to 'use the task tool to run steps in subagents' (instead of just 'use subagents') so possibly the name of the tool confuses models?