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 override is worse than none. A single quick lookup doesn't need an agent at all.
Workflow Scripts
For agent(prompt, {model, effort}):
- Mechanical stages (fetch, extract, classify): small or mid tier,
effort: 'low'
- Judge/verify stages: top tier, higher effort — that's where hallucinations get caught
- Map→reduce→judge pattern: cheap finders, expensive judges
Instance: testing-with-synthetic-users runs its host top-tier, personas mid-tier.
Capped Top Tier
If your plan caps top-tier usage: that quota belongs to the main loop — orchestration, triage, final synthesis. Dispatch the table's "Top" work one tier down instead (judges, design, heavy implementation); the main loop reviewing their output IS the top-tier quality pass. The top model thinks; agents do. One tier down is typically half the price and still outclasses mid-tier finders — "judges must outclass finders" holds.
Personalization
Bind routing to agent types: define a custom agent in ~/.claude/agents/<name>.md with model: pinned in its frontmatter (e.g. a taste-designer pinned top-tier). For recurring task types the table doesn't cover, fork this skill and add rows.
Common Mistakes
| Mistake |
Fix |
| "This research is complex, it needs the strongest model" |
Complexity ≠ judgment. Research is mid tier; spend top tier verifying the findings instead |
| Routing design/taste work down-tier "to save cost" |
Taste degrades worst on cheap models — always top tier |
| Paying top or mid tier for sweeps, inventories, bulk extraction |
Mechanical bulk is small tier — "mid tier to be safe" is the most common silent overspend |
| Guessing model id strings |
Use your platform's exact ids (Claude Code: "sonnet", "opus", "haiku", "fable") |
Overriding model on a custom agent that pins one |
Its frontmatter already routes it — leave it alone |
1---2name: choosing-agent-models3description: Use when about to dispatch subagents or parallel agents — the user asks to fan out, split work across multiple agents, or delegate research, implementation, or review to workers — when choosing a model or effort override for any agent, or when authoring workflow scripts with agent() calls4---56# Choosing Agent Models78## Overview910Route by **task nature**, not task importance. Three tiers:1112- **Top tier** — strongest available (e.g. fable, else opus): taste and judgment — design, critique, architecture decisions, adversarial verification, multi-agent synthesis13- **Mid tier** (e.g. sonnet): research, exploration, standard implementation, content drafting14- **Small tier** (e.g. haiku): mechanical bulk — inventory, format conversion, bulk extraction/classification (heuristic: many similar items, no per-item judgment)1516Adjacent 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.1718## Quick Reference1920| Task | Tier |21|---|---|22| Visual/frontend design, redesign, brand work | Top — via a pinned design agent (see Personalization) |23| Design critique, architecture decisions | Top |24| Adversarial verification, judging, final synthesis | Top — judges must outclass finders |25| Web/market research, competitor analysis | Mid — NOT top |26| Multi-file code exploration, standard implementation | Mid |27| Content drafting | Mid |28| Repo/log sweeps — "map X, report a structured list" | Small — a report output doesn't make a sweep judgment work |29| Format conversion, bulk extraction/classification | Small |30| Unsure, mixed-nature, or novel | **Omit `model`** — inherit the session model |3132## When NOT to Override3334Only pass `model` when the task clearly fits a row above — a wrong override is worse than none. A single quick lookup doesn't need an agent at all.3536## Workflow Scripts3738For `agent(prompt, {model, effort})`:3940- Mechanical stages (fetch, extract, classify): small or mid tier, `effort: 'low'`41- Judge/verify stages: top tier, higher effort — that's where hallucinations get caught42- Map→reduce→judge pattern: cheap finders, expensive judges4344Instance: testing-with-synthetic-users runs its host top-tier, personas mid-tier.4546## Capped Top Tier4748If your plan caps top-tier usage: that quota belongs to the main loop — orchestration, triage, final synthesis. Dispatch the table's "Top" work one tier down instead (judges, design, heavy implementation); the main loop reviewing their output IS the top-tier quality pass. **The top model thinks; agents do.** One tier down is typically half the price and still outclasses mid-tier finders — "judges must outclass finders" holds.4950## Personalization5152Bind routing to agent types: define a custom agent in `~/.claude/agents/<name>.md` with `model:` pinned in its frontmatter (e.g. a `taste-designer` pinned top-tier). For recurring task types the table doesn't cover, fork this skill and add rows.5354## Common Mistakes5556| Mistake | Fix |57|---|---|58| "This research is complex, it needs the strongest model" | Complexity ≠ judgment. Research is mid tier; spend top tier verifying the findings instead |59| Routing design/taste work down-tier "to save cost" | Taste degrades worst on cheap models — always top tier |60| Paying top or mid tier for sweeps, inventories, bulk extraction | Mechanical bulk is small tier — "mid tier to be safe" is the most common silent overspend |61| Guessing model id strings | Use your platform's exact ids (Claude Code: `"sonnet"`, `"opus"`, `"haiku"`, `"fable"`) |62| Overriding `model` on a custom agent that pins one | Its frontmatter already routes it — leave it alone |
Kelonwai