A classifier agent reads an incoming context packet and dispatches it to the appropriate specialist. The router owns no domain knowledge — it owns the decision of who handles what.
A B2B SaaS company receives 2,000 support tickets per day: billing disputes, API errors, feature requests, security reports, and account cancellations. Each type requires a completely different handling path. Billing disputes go to an agent with CRM and payment system access. API errors go to an agent with log access and engineering escalation. Security reports must reach a human immediately.
Running all 2,000 tickets through a single capable model costs 10x more than routing to specialized handlers. But routing accuracy is everything: a security report misclassified as a feature request is a breach waiting to happen. The router's job is to make this decision correctly, cheaply, and at scale.
| Node | What it does | Input | Output |
|---|---|---|---|
| Ticket Classifier | Reads ticket text and assigns a category label. Does not resolve anything — only categorizes. | Raw ticket text + user metadata | Category label + confidence score |
| Billing Agent | Resolves billing disputes, refund requests, invoice questions. Has CRM and Stripe access. | Ticket + billing history | Resolution + updated CRM record |
| Technical Agent | Diagnoses API errors, latency issues, integration failures. Has log access and runbook library. | Ticket + error logs + API keys | Root cause + remediation steps |
| Security Escalation | Immediately pages on-call human. No AI resolution — direct human handoff only. | Ticket classified as security | PagerDuty alert + ticket frozen |
| Feature Request Logger | Extracts feature request, deduplicates against existing backlog, creates Linear ticket. | Ticket + product backlog | Linear ticket + acknowledgment |
| Metric | Signal |
|---|---|
| Routing accuracy | Primary classifier health signal — benchmark against human labeling monthly |
| Class distribution drift | Leading indicator of accuracy degradation — alert on >15% weekly shift |
| Cost per routed ticket | Classifier cost vs. specialist cost savings — ROI of routing decision |
| Fallback branch rate | Proxy for unhandled variety — rising rate signals new input class emerging |
| Origin of Value | Where it appears | How it is captured |
|---|---|---|
| Value Transfer | Classifier → Specialist edge | The router moves each ticket to the handler best positioned to extract value from it. Correct routing is the primary value act — the specialists only receive value if routing is correct. |
| Governance | Classifier node | The routing decision determines which specialist gets authority over the ticket. A security ticket routed to the Billing Agent is a governance failure with immediate, irreversible consequences. |
| Risk Exposure | Every misclassification | Misrouting destroys value immediately. Class distribution drift degrades classifier accuracy silently over time — a risk that compounds with no visible signal until a major misroute surfaces. |
| Conditional Action | Classifier compute cost | If the classifier is a large model, its cost must be weighed against downstream savings. A GPT-4o classifier routing to GPT-4o mini specialists is only profitable if routing accuracy is high enough. |
VCM analog: Access Token. The router grants or denies access to specialist agents. Like an access token controlling which resources a principal can reach, the classifier determines which specialist receives the work. Misrouting = revocation of the right specialist's access.
A security report is classified as a feature request with 0.87 confidence. The Feature Request Logger creates a Linear ticket and sends an acknowledgment. The security vulnerability is now in a product backlog, not in incident response. The output looks valid because the specialist did its job correctly on the wrong input. Detection requires human spot-check of high-stakes categories — automated detection is only possible if the downstream specialist validates the ticket type matches its expected class.
The classifier was trained on ticket data from 12 months ago. API error tickets have shifted in vocabulary — users now say "webhook timeout" instead of "integration failure." The classifier confidence for Technical tickets drops from 0.91 to 0.74 over 6 months with no alert. Fix: monitor routing distribution weekly. A sudden shift in category proportions is the leading indicator of classifier degradation.
A new input type appears — enterprise contract renewal requests — that the classifier was not trained on. It routes to Billing Agent (closest match). Billing Agent cannot resolve contract terms; ticket stalls. The fallback route must be a first-class citizen: an explicit "unknown" category that triggers human review, not a default to the most similar known class.
| Variant | Modification | When to use |
|---|---|---|
| Confidence-Gated Router | Routes only if confidence exceeds threshold; else routes to human review | High-stakes routing where uncertain classification is costly (security, legal, compliance) |
| Cascading Router | Multiple classifiers in sequence, each narrowing the class | Large specialist space — single classifier would have too many classes to be accurate |
| Cost-Aware Router | Routing decision includes estimated task complexity; simple → cheap model, complex → capable model | Heterogeneous model fleet where cost optimization is a primary goal |
| Pattern | Relationship |
|---|---|
| 10.11 Pipeline | Each router branch typically runs a pipeline — the router selects which pipeline to activate |
| 20.23 Orchestrator-Workers | Dynamic routing variant — Orchestrator routes at runtime based on task decomposition, not pre-defined classes |
The router is a strategic asset. The firm that owns the routing layer owns dispatching logic for the entire system. Router quality determines whether acquired specialist agents receive well-formed work — a bad router makes every specialist underperform.
A classifier trained on proprietary historical data — customer request vocabulary, domain-specific ticket categories, behavioral signals accumulated over years — is hard to replicate. The routing IP is often more defensible than the specialist models behind it.
Due diligence question: can the firm report routing accuracy by category, week over week? If not, they cannot detect drift and do not know whether their specialists are being used correctly.