Developer advocates for graph-based engineering over traditional coding for scalable AI systems

A developer with years of experience building LLM features argues that traditional intent detection and nested conditional code creates unscalable systems. They state the core problem is that agent flows exist only in code and stale diagrams, not as observable artifacts. Their proposed solution is declaring system flows as graphs rather than writing control flow code. This approach, developed years before it became a named practice, aims to address traceability, testing, and change management in production AI systems. The developer bases this argument on lessons from real financial-domain products built with real teams and users.