auto-agent-eco

/auto-agent-eco

Spawn subagents one tier below the model doing the orchestrating. The expensive model decides, routes and merges; cheaper models do the legwork.

Usage: /auto-agent-eco — the agent count is yours to judge. /auto-agent-eco 6 — a hard ceiling of six subagents.

The Rule

Read the model this session is running as, then cap every subagent strictly below it:

Orchestrator	May spawn	Never
Fable	Opus, Sonnet, Haiku	Fable
Opus	Opus, Sonnet, Haiku	above Opus
Sonnet	Sonnet, Haiku	above Sonnet
Haiku	Haiku	anything above

Note the asymmetry: from Fable, the ceiling drops a full tier — that is the entire point of eco mode, and it is where the saving comes from. From Opus and below, "strictly below" would leave nothing useful to route to on hard subtasks, so same-tier is permitted and the discipline becomes routing rather than capping. Do not spawn a peer out of habit; spawn one when the subtask genuinely needs that tier.

If the session's model cannot be determined, assume the tier below your best guess and say which assumption you made. Guessing high spends the user's money on an assumption they never agreed to.

Where The Saving Actually Comes From

Not from making everything cheap. From splitting the work by kind:

• The orchestrator keeps the judgment — deciding what the subtasks are, spotting when a result is wrong, reconciling contradictions, and writing the synthesis. That work is dense, short, and the worst possible thing to hand to a cheap model.
• Subagents get the volume — reading files, tracing call sites, running greps and builds, drafting mechanical edits, gathering evidence. That work is long, parallel, and mostly legwork.

A pass that hands twelve files to a cheap model and one paragraph of conclusions to the expensive one costs a fraction of a uniform fan-out and loses very little. A pass that hands the conclusions to the cheap model saves nothing worth having, because the orchestrator then re-does the thinking.