Conditional Routing

"Don't ask one agent to handle every case. Route the request to the agent that knows this case."

Context

Incoming requests vary in nature. A customer support system receives billing questions, product inquiries, technical issues, and complaint escalations. Each type benefits from a different agent with different skills, different tool access, and different constraints. One general-purpose agent handles all of them poorly.

Problem

A single agent handling all request types either has too-broad tool access (security risk), too-generic constraints (quality risk), or too-complex specifications (maintenance risk). But naively routing "by topic" produces brittle classifiers that break on ambiguous inputs.

Forces

  • Specialization vs. routing complexity. Specialized agents are more accurate. But routing to the right specialist requires classification that is itself an agent task.
  • Deterministic routing vs. fuzzy inputs. Rule-based routing (keyword matching) is fast but brittle. Agent-based classification (Advisor archetype) is flexible but adds latency and may misclassify.
  • Route coverage vs. unknown inputs. Every routing path must be specified. But inputs occasionally fall outside known categories. An unhandled route produces a silent failure or a catch-all that undermines specialization.

The Solution

Implement routing as a classifier agent (Advisor archetype) whose only job is to analyze the input and produce a routing decision — not content.

Routing structure:

  1. The classifier agent receives the input and produces a structured routing decision: { "route": "billing", "confidence": 0.92, "reasoning": "customer mentions invoice and payment" }.
  2. Routes are declared in the pipeline spec — each route names the destination agent, its archetype, and its spec.
  3. A default route handles unknown inputs. Typically: escalate to human with the classifier's analysis, or return to the user asking for clarification.
  4. Low-confidence classifications trigger escalation rather than low-confidence routing. A confidence threshold is declared in the spec.

Routing criteria are declared, not inferred. The classifier isn't making a creative decision — it's checking the input against declared routing criteria specified in its skill file.

Resulting Context

  • Each destination agent is specialized. It has narrow constraints, focused tool access, and domain-specific skills — producing higher quality than a generalist.
  • Routing is auditable. The classifier's decision, confidence, and reasoning are logged, making misroutes diagnosable.
  • New routes can be added without modifying existing agents. Adding a new request type means adding a route and a new specialized agent, not modifying the generalist.

Therefore

Route varied inputs through a classifier agent (Advisor archetype) that produces structured routing decisions. Each route leads to a specialized agent with focused constraints and tools. Declare all routes in the pipeline spec, including a default for unknown inputs.

Connections

  • Sequential Pipeline — routing is the first stage in a pipeline, directing the request to the appropriate sub-pipeline
  • Parallel Fan-Out — some inputs may need to be processed by multiple agents simultaneously
  • Escalation Chain — low-confidence or unknown classifications escalate rather than route
  • The Five Archetypes — the classifier is Advisor; destination agents are typed by their archetype