when fable got re-enabled and then Sol dropped i’ve noticed that subagents will often use the same model as parent; this is less than ideal, because of the cost, speed, etc
i’ve experimented with custom instructions to steer subagents towards smaller and cheaper models (mostly Composer, but also Luna/Terra) with mixed results, so ended up dropping these rules and manually instructed the models to run Composer
now, since yesterday or so i’ve noticed that most of the subagents are actually Composers, even without explicit instructions; which is great
has anything changed or is it just the luck of the draw? can you publish the prompts you use to steer agents towards cheaper models in subagents?
There isn’t a separate “prompt” you can publish here. Model selection for subagents works differently. In short:
Regular Task subagents inherit the parent agent’s model by default, unless a model is set explicitly.
The explore subagent has its own default (the Composer family). You can override it in Settings.
For custom subagents, the model is fixed via model: in the frontmatter, or the parent can pass a specific model directly in the Task call.
About “did anything change over the last couple of days”, I can’t confirm a specific change in that exact area. Subagents leaning toward cheaper models (Composer) is already established behavior, so most likely you just stopped overriding it with your rules, and the usual model rotation also happened (they brought fable back, removed Sol).
If you need strict control, not “whatever happens”, pin the model explicitly via frontmatter or Settings. That’s more reliable than relying on the default. More on subagent config: Subagents | Cursor Docs
I also want to plug the workflows/subagents implementation in Claude Code (first create a javascript file it then uses to orchestrate the agents). Probably something the team is already thinking about.