routing

Define a small set of routes with clear criteria, then use a classifier step that outputs a label plus confidence. Send each labeled request to its specialist prompt, model, or tool chain, and include a fallback path for low-confidence cases.

Use routing when requests fall into clear categories that need different prompts, tools, or models, and a cheap classifier can pick the path reliably. Skip it when categories overlap heavily or a generalist already performs well.

Common routing axes include intent type, task complexity, safety level, domain, required tool access, and response format. Choose axes where categories have low overlap so the classifier can decide confidently.

Misroutes usually come from overlapping route definitions, too many routes, or no fallback for ambiguous inputs. Save misroute examples, refine route criteria, and track per-route accuracy metrics to find weak spots.

Routing fails when categories are fuzzy, misroute costs are high, or there are too few examples to define routes. It also adds maintenance burden since each route needs its own tested prompt and periodic review.