Agent Dispatch — AI skill for Claude Code

Agent Dispatch — which model runs each subagent

[](LICENSE) [](SKILL.md) [](README.es.md)

A free, standalone skill that stops every subagent from running on your biggest model.

By default, Claude Code hands every subagent the model you have selected for the session. So when a task fans out into nine agents, all nine run on Fable 5.1 (or Opus, or whatever you picked) — the four that translate a `messages.json` file cost the same per token as the one that designs the architecture. This skill makes the split explicit before anything launches: **the model is picked by the type of task, never by the importance of the project.**


The rule in one table

Type of task	Examples	Model · effort
Mechanical	translations of fixed strings, listing, counting, measuring, running a script and reporting the output, checking that a manifest parses	small (e.g. Haiku) · low
Bounded analysis	summarizing a module, mapping dependencies, writing tests from a clear spec, one mechanical change across many files	mid (e.g. Sonnet) · medium
Judgment	architecture, writing the code that decides how something works, choosing between options, the final review, anything that ships	the session's model (e.g. Fable 5.1, Opus) · high

Plus four guardrails the skill enforces: the final reviewer never gets cheapened; a small agent that fails moves the task up a tier instead of being retried; fewer agents beat more agents; and every run **declares its split before launching and reports the measured spend after** — or says "not measured". No invented savings.

Worked example: a