Tip: Give Agents Judgment, Not Scripts — Delegate the Small Stuff to Lower-Tier Subagents

Simon Willison·medium signal

In "Judgement" (July 3), Willison argues for letting capable models exercise independent judgment instead of over-prescribing steps, while routing smaller, well-scoped tasks to lower-tier models via subagents. Concretely: reserve Opus-class reasoning for the plan and the hard calls, and fan out mechanical subtasks to cheaper models — the same heterogeneous-routing pattern, but stated as a prompt-authoring discipline.