choosing-agent-modelslisted

# Choosing Agent Models ## Overview Route by **task nature**, not task importance. Three tiers: - **Top tier** — strongest available (e.g. fable, else opus): taste and judgment — design, critique, architecture decisions, adversarial verification, multi-agent synthesis - **Mid tier** (e.g. sonnet): research, exploration, standard implementation, content drafting - **Small tier** (e.g. haiku): mechanical bulk — inventory, format conversion, bulk extraction/classification (heuristic: many similar items, no per-item judgment) Adjacent tiers differ several-fold in cost. Routing an inventory sweep down a tier is free money; sending a design critique down is where cheap models fail worst. ## Quick Reference | Task | Tier | |---|---| | Visual/frontend design, redesign, brand work | Top — via a pinned design agent (see Personalization) | | Design critique, architecture decisions | Top | | Adversarial verification, judging, final synthesis | Top — judges must outclass finders | | Web/market research, competitor analysis | Mid — NOT top | | Multi-file code exploration, standard implementation | Mid | | Content drafting | Mid | | Repo/log sweeps — "map X, report a structured list" | Small — a report output doesn't make a sweep judgment work | | Format conversion, bulk extraction/classification | Small | | Unsure, mixed-nature, or novel | **Omit `model`** — inherit the session model | ## When NOT to Override Only pass `model` when the task clearly fits a row above — a wrong overri