Dan M

We tracked code quality metrics across 14 engineering teams using AI coding agents and found that while velocity increased by 40%, architectural coherence declined by 28% and cross-module defects tripled. Teams are measuring the wrong thing.

Every AI agent session starts from zero. The architectural decisions from six months ago, the module boundaries drawn for specific reasons, the bug fixed three times in three different ways. We examined how knowledge loss compounds across agent sessions and what it costs.

The dominant workflow for AI coding agents is prompt-and-pray: describe what you want, hope the output is correct. We studied 11 engineering teams and found that teams using structured design canvases before agent execution produced 62% fewer cross-module defects and 3x faster review cycles.

Code review was designed for human-speed output. When agents produce ten times the volume, manual review becomes theatre. We found that reviewers catch 23% of defects in agent-generated code versus 61% in human-written code, and the gap widens as volume increases.

The prevailing assumption is that autonomous AI agents are the future of enterprise automation. We argue that auditable, deterministic pipelines with human validation produce better outcomes, and explain why the distinction matters.