Why multi-step AI workflows are overhyped

In this post, Utkarsh Kanwat put my perception about multi-step AI workflows into words. While AI is part of my coding workflow, I’ve been struggling with multi-step, agentic workflows. For one-step coding tasks I get good results with models like Claude Sonnet 4.5 or Haiku 4.5, but for anything more complex that requires multiple steps, the perceived value I get from AI degrades quickly.

Utkarsh weighs in on the hype around multi-step AI agents. He argues that:

• error rates compound across multi-step workflows, so long chains become unreliable;
• conversational context windows create quadratic token costs, making long interactions economically unsustainable; and
• agent tools and integrations are hard to engineer because they often require human confirmation gates.