What I Learned While Writing 2 Million Lines of Code with AI

A veteran CTO says he and AI produced over 2 million lines of code across services, agents, and apps, but much of it only worked in demos and broke in real-world runtime conditions.

He argues the real missing piece is context: the purpose, constraints, and operating environment of the software, not just feature specs or implementation steps.

Adding that context made Claude Code’s suggestions noticeably better for his solo workflow, showing how much context engineering can matter.

The story warns that AI coding gains can be misleading if teams measure success by output volume or polished demos instead of resilient production behavior.

It also reinforces concerns about “agent theater” and “agent washing,” suggesting better context may matter more than switching models.