AI systems expose gaps in documentation, testing, and implicit developer knowledge that were manageable when humans navigated codebases but become live production risks when AI does

AI finds production vulnerabilities by navigating code paths that experienced developers unconsciously avoid. The article describes undocumented callback hooks, stale test suites, and implicit knowledge about third-party service load limits as problems that "worked great for humans" but cause AI to trigger deadlocks and outages because it lacks institutional memory.

Developer tools built for human consumption (like RSpec, Rubocop, Minitest) force AI to repeatedly parse text output and re-run commands to diagnose failures, when structured JSON output would let it identify issues in one pass. The author has begun requiring all AI tools to output JSON and using wrapper functions to guarantee structured output.

The solution is framing AI failures as system design failures, not agent mistakes—asking "what failed in this system that allowed this error" rather than blaming the AI. This mirrors how good engineers respond to human-caused production incidents: learn, adapt, and build guardrails (testing, documentation, deterministic validation, static typing).