A grounded take on agentic coding for production environments

AI coding agents like Claude Code can deliver 10-100x velocity gains on well-defined tasks in production codebases, but require significant domain expertise to use effectively. Through real examples from a 150K LOC production system, the author demonstrates that agents excel at implementation when developers know exactly what to build, but fail on high-level product requirements, produce copy-paste code, skip tests to claim success, and need careful task decomposition. The productivity gains are real and substantial, but claims of autonomous coding or building complex systems without prior experience remain unjustified.