Subagents and the scout tier

When scout subagents are on, the agent hands exploration to a cheaper model: reading files, searching for symbols, scanning logs.

On by default. Settings → Agent → Scout subagents.

Why this works

Exploration is most of an agent's token spend and almost none of its difficulty. Finding which file defines a function is not the hard part — deciding what to change about it is.

Running that search on a frontier model is paying premium rates for work a small model does perfectly well.

What scouts return

A short structured briefing, not raw file contents. A scout that read four files comes back with what each one is and whether it matters — not four files pasted into the main conversation.

That is the second benefit: the main agent's context stays small, so it can work longer before compaction and reason over what actually matters.

What they never do

Scouts do not write code, edit files or run destructive commands. They read, search and report. Anything that changes your project is done by the main agent, which you are already reviewing.

What it costs

Usually less than not using them. Two effects compound: exploration moves to a cheaper tier, and the main model's context stays smaller — and input tokens are charged on every turn.

When to turn it off

If you are debugging the agent's behaviour and want a single unbroken transcript, or working on something where the exploration itself needs the strongest available model — a subtle bug across unfamiliar code, say.

See also

• Agent and scout tiers
• Context and compaction