router

flowchart TD
    Input([Task input]) --> Router[Router agent]
    Router --> C{Category}
    C -- BUG_FIX --> B[Bug-fix agent]
    C -- FEATURE --> F[Feature agent]
    C -- REFACTOR --> R[Refactor agent]
    C -- SECURITY --> S[Security agent]
    B & F & R & S --> Done([Result])

A classifier agent analyzes the input and dispatches to the appropriate specialist agent. The routing decision is made once, upfront — no specialist ever sees tasks outside its domain.

How it works

1. The router agent classifies the input into a known category (e.g. BUG_FIX, FEATURE, REFACTOR, SECURITY).
2. The category determines which specialist agent receives the task.
3. The specialist processes the task with a focused system prompt.
4. Effort is allocated per role: the router runs at low effort with a small max_tokens (one word out); the specialist gets the budget at medium. Same model, one cache namespace — measure this before reaching for a cheaper classifier model.

When to use

• Pipelines where input type varies significantly and a single generalist agent produces worse results than a specialist.
• Systems where you want each agent’s context to stay narrow and auditable.

When not to use

• Tasks that span multiple categories at once (e.g. a bug fix that also requires a security patch) — the router will misclassify or you’ll need a second pass.
• When the category set is unstable — adding new categories requires rewriting the router prompt.

Trade-offs

Failure modes

• Misclassification — the router picks the wrong category; the specialist produces a confident but irrelevant result.
• Unknown category — input doesn’t fit any known category; the router hallucinates a label or falls through to a default.