Routing Agents

Routing agents classify an incoming task or message and direct it to the most appropriate resource — another agent, a tool, a human, or a workflow. They are the traffic controllers of multi-agent systems.

When Routing Matters

As agent systems grow in complexity, not every request should go to the same model or tool:

Cost efficiency — a complex coding question deserves a reasoning model; a simple FAQ deserves a fast, cheap model
Specialisation — different agents may be optimised for different domains (legal, technical, creative)
Safety — sensitive topics may require human escalation before any agent handles them

Routing Approaches

Approach	Mechanism	Trade-offs
LLM classifier	Prompt an LLM to classify the intent	Flexible, adds latency and cost
Semantic similarity	Embed the query, find closest category	Fast, requires good category definitions
Rule-based	Keyword or regex matching	Fastest, brittle to edge cases
Hybrid	Rules first, LLM fallback	Balanced cost and accuracy

OpenAI Agents SDK: Handoffs

The OpenAI Agents SDK (March 2025) formalised routing through its Handoffs primitive — a structured way for one agent to transfer task execution to another, with full context passing. This replaced ad-hoc prompt-based routing patterns.

Source

OpenAI Agents SDK: Handoffs, March 2025

Key Considerations

Routing decisions should be logged and auditable — when a task is misrouted, you need to diagnose why
Include a fallback handler for queries that don't match any route — never silently drop messages
Test routing accuracy separately from downstream task accuracy; errors compound