When we started building our multi-brand AI search assistant, the industry buzz pointed us toward autonomous agents. We evaluated the popular frameworks — LangGraph, CrewAI, AutoGen, and others — each promising flexible, self-directing AI systems. But as we dug deeper, concerns emerged: unpredictable latency as agents decided their own paths, token costs multiplying with each reasoning loop, and debugging sessions that felt like archaeology.

We stepped back and asked: what do we actually need? Our assistant handles product search, recommendations, and customer questions across multiple retail brands. The tasks are complex but well-defined. Users expect fast, consistent responses. We need to test, monitor, and explain system behavior.

The answer wasn't agents — it was smart routing with simple flows. We built our system on Genkit with a single architectural principle: classify intent once, then execute deterministic flows. One LLM call decides the path (product search, clarification, comparison, off-topic). From there, specialized flows handle RAG retrieval, filter extraction, and streaming response generation — all predictable, all testable.

The results speak for themselves: sub-second routing decisions, consistent response times, straightforward debugging, and a codebase that new team members can understand in days, not weeks. We now serve multiple enterprise clients across retail and cosmetics, with the same architecture handling different brands, catalogs, and deployment targets.

This talk shares our evaluation journey, the "smart routing, simple flows" architecture we landed on, concrete implementation patterns with Genkit, and honest lessons about when this approach works — and when you might actually need agents.