Concept · Delegation · 3 min read

Tasks and delegation

Most token spend is mechanical work done by an expensive model. Delegation is Kaptain's structural answer: the strong model plans and judges; suitable work routes to a local worker whose tokens cost you nothing.

The split that saves money

Summarize a file, extract fields, draft boilerplate, digest a log: none of it needs your best model, and all of it burns your best model's budget when routing is manual. Kaptain gives the split a mechanism instead of a habit.

How the delegate flow works

The delegate tool hands one bounded task, with named context files, to the configured local worker; delegate_batch fans out independent items and writes each result to disk so only stats return to the caller's context. The worker model is one Settings choice, applied live — and over the mesh it can be a stronger box's model instead of a weaker local one.

The quality gate

Delegated output passes a gate before anyone downstream trusts it; a failed gate is surfaced as a failure, never silently substituted with worse output. That's a standing rule across Kaptain: degraded modes are labelled, not hidden.

What to delegate — and what never

Delegate work that is mechanical, bounded, and reviewable. Keep ambiguous reasoning, architectural judgment, and final synthesis on the strong route. Scheduled kapsules compose with this naturally: a cheap watcher runs hourly and only escalates to the strong model when something actually changed.

Agent-readable version: index.md