Every dispatch names a model. Omitting the field is not a neutral default — it silently runs the subagent on the parent's model, so a fan-out of file reads burns orchestrator-grade spend and a security review runs on whatever happened to be driving the session.

The goal is never the best model. It is the cheapest model that clears the bar — the least capable model that still finishes the task in one pass. Overshooting wastes spend on every dispatch; undershooting costs more than it saves, because a model that loops three times and still needs correcting is more expensive than one pass on the tier above.

Selection runs on four dimensions, scored 1-10, higher is better:

Speed — latency and throughput.
Taste — output craft: prose, coding, UI/UX, restraint.
Intelligence — reasoning depth, long-horizon reliability, tool use, completion rate.
Cost — efficiency of total spend to finish the task, not price per token.

Classify

Match the task to a role. The role fixes a leading dimension and a bar — the minimum score on that dimension the task demands. Classify on what the task actually demands, not on what it is called.

picking-models

Classify