Do you split your AI subagents into tiny narrow roles or broad ones?
by•
I've been managing a dozen narrow agent configs where each file limits tools and picks a specific model like Opus for architecture planning and Sonnet for basic coding.
But assigning heavy models to review tasks gets expensive fast and juggling too many sub-roles makes the delegation chain messy.
If you use subagents in your dev workflow, do you prefer strict tool boundaries for every single task, or just two or three general-purpose agents?
8 views
Replies
Our setup works the other way. Most rules live in one shared CLAUDE.md plus a CLAUDE.md per project, and 78 hooks block the things a prompt can't be trusted with. I keep the model's judgement to a minimum and push the rest into scripts, and when it does have to judge, a different model checks it. Cost isn't really the issue for me since I'm on a fixed Max plan.
i for sure break up a task into focused stages and assign a specific model to each stage. for example, when doing pull request code reviews, i usually have codex astra xhigh analyze the code through a Correctness lens, fable through a Quality lens, and then just let antigravity or something cheaper like sol capture before/after screenshots for Evidence. when compared with code reviews done by other devs on the team, mine really stand out when they catch issues while others have already supposedly approved the PR.