Model selection & routing
Match the model to the task so the economical choice is the default choice — encoded into the agent config defaults, not left to per-session discretion.
Routing
| Task type | Model tier |
|---|---|
| Boilerplate, formatting, simple tests, mechanical refactors | fast / cheap |
| Feature work, standard review | mid |
| Architecture, gnarly debugging, security-sensitive reasoning | frontier |
Caching and context discipline (load only what's needed — see repo context & memory) apply at every tier; wasted context is wasted spend.
Change management
AI platforms churn monthly. When a provider ships, updates or deprecates a model:
- Re-run the golden-task harness.
- Compare cost/quality against the incumbent.
- Pin vs migrate decision, recorded as an ADR.
- Update the config default; issue a squad migration note.
Without this routine, every model change is a fire drill. With it, it's a scored decision. Model IDs and tiers are pinned in config so a silent upstream change can't alter behaviour unnoticed.
Standards referenced: ITIL 4 change enablement, FinOps.