Why AI Agent Routers Break in Production and What They Should Do Instead

A software developer reflects on the widespread use of keyword-based intent routing in AI agent platforms, a pattern that routes user requests to specialist agents based on simple pattern matching. The approach works well in demos but breaks down in production due to three core failures: ambiguity in underspecified requests, semantic collisions at scale, and the inability to audit past routing decisions. After hitting all three problems simultaneously during a production incident, the developer shifted focus from what a router can do to what it owes every caller. This led to a framework built around five obligations, starting with generating a structured, typed fingerprint of each request before any routing decision is made. The fingerprint captures intent class, domain tags, classification confidence, and ambiguity signals, serving as a prerequisite for accountable and reliable routing.