AI coding agents need evidence-first review, not just cheaper routing
The article argues that while AI coding agents are becoming cheaper, the real productivity bottleneck is not cost but the need for evidence-first review processes to ensure code quality and reliability, rather than simply optimizing for cheaper routing.
Background
- The article argues that the real cost of AI coding agents isn't "routing" (choosing which model to use for each task), but rather the cost of verifying that AI-generated code is correct and safe. This "evidence-first review" cost is often ignored in discussions about making AI coding cheaper.
- "Routing" means dynamically picking between cheap/small LLMs and expensive/powerful ones for different steps in a coding workflow. Much current optimization focuses on this to reduce token spend.
- The piece pushes back: even with perfect routing, every line of AI code still needs testing, review, and security checks — and that human/automated verification is the true bottleneck.
- Published on undes.app, a blog by a developer focused on software engineering productivity and AI-assisted coding. Written for readers already familiar with LLM cost debates and AI coding tools.