Routing

Anthropic canonical agent pattern. Classify first. Then send work to best path.

What It Is

Use a router to choose among specialists. Specialist can be prompt, model, tool chain, or agent. Goal: better fit per request.

When to Use

• Requests fall into clear categories
• Different categories need different prompts or tools
• Cheap classifier can pick path well
• Specialist path beats one general path
• Need cost or latency control by request type

When Not to Use

• Categories fuzzy or overlapping
• Generalist path already works
• Misroute cost high
• Too few examples to define routes
• Router logic harder than task

Core Flow

input
  → classify
  → pick route
    → specialist A
    → specialist B
    → specialist C
  → return result

Simple Implementation Outline

1. Define route set.
2. Write route criteria.
3. Start with few routes.
4. Make router output label + confidence.
5. Add fallback path.
6. Track route accuracy.
7. Review misroutes. Refine criteria.

Good Routing Axes

• Intent type
• Task complexity
• Safety level
• Domain
• Required tool access
• Response format

Failure Modes

• Too many routes. Hard to maintain.
• Route definitions overlap.
• No fallback for low confidence.
• Router prompt leaks to specialists.
• Uneven traffic. Some routes rot.
• No eval set. Misroutes stay hidden.

Practical Checklist

• Clear route taxonomy
• Low-overlap route definitions
• Confidence or abstain option
• Fallback path exists
• Per-route prompts tested
• Misroute examples saved
• Metrics by route: quality, cost, latency
• Periodic route review

Decision Rule

Use routing when specialization wins and classification is cheap enough. If cases blur together, keep one path.