Delegate implementation to a lower-tier model in a subagent; keep judgment work in the top model

Simon Willison's Weblog·medium signal

A practitioner pattern gaining traction: for coding tasks, spawn a subagent running a cheaper model (mid-tier for substantive edits, smallest tier for mechanical/trivial changes) with a self-contained prompt, then review its result in the main loop before committing. Design, auditing, data synthesis, and judgment-heavy reasoning stay on the top model. This routes the bulk of token spend to cheap models without giving up the expensive model's taste on the decisions that matter.