AllanWessels Agent Model Selection
When to use
- Any time you dispatch a sub-agent via the Agent tool.
- When the task mix includes both high-reasoning (architecture review, complex debugging) and low-reasoning (run a test suite, apply a grep, format a file) work.
When NOT to use
- You are the sub-agent (model selection is for the dispatcher).
How to apply
Pass model: explicitly in every Agent tool call. Do not rely on inheritance
from the parent session:
model: "sonnet" # fast, cheap — use for: test runs, file reads, curl checks,
# mechanical code changes, grep-and-report tasks
model: "opus" # slow, expensive — use for: architecture review, deep
# reasoning over a large codebase, novel algorithm design,
# judge agent (when correctness is load-bearing)
Decision table
| Task type | Model |
|---|---|
| Run pytest / vitest / Playwright and report results | sonnet |
| Apply a targeted code fix from a spec | sonnet |
| Execute a pre-handoff checklist (curl + ls) | sonnet |
| Audit a test suite for structural gaps | sonnet |
| Design a new subsystem architecture | opus |
| Deep code review over 10+ files | opus |
| Judge / evaluator agent (correctness is load-bearing) | opus or sonnet pinned |
| Multi-file refactor from a spec | sonnet |
The judge exception
If the model is acting as a stable evaluator / judge whose output is trusted as ground truth (e.g. a RAG judge agent), do not downgrade it mid-run for cost reasons. Inconsistent judge model = inconsistent scores = the optimization loop optimizes noise. Pin the judge model in config and enforce it.
Why this works
Sub-agents inherit the parent session's default model unless overridden. In a
session where the parent is using a strong model for architecture work, every
dispatched sub-agent (including "just run ls and report back") also runs on
that model. At scale — dozens of parallel verification agents — this cost
multiplies significantly. Explicit selection ensures each agent uses the
minimum capable model for its task.
Anti-patterns to avoid
- Omitting
model:and assuming inheritance is intentional — it is accidental. Always be explicit. - Downgrading a judge or evaluator agent to save cost — the judge's reliability is a system invariant, not an optimization target.
- Using a strong model for a mechanical task "just to be safe" — this is false safety. A test runner either passes or fails; a stronger model does not change the test results.
Cross-links
- [[parallel-fanout-verification]] — where most sub-agent dispatches happen; all test-runner agents should use sonnet