"Code Quality Doesn't Matter Anymore" Is Wrong. I Measured It

Everyone says code quality doesn't matter anymore: AI agents write the code, humans don't need to read it. I wanted an engineer's answer, so I built an experiment. Same backend, built twice: one app with everything in a single 9,000-line file, one app with a layered structure and a small resource kit. Same API, same 45 resources, both pass the same 1,355 tests. Then two AI agents shipped the same two releases into both apps, one after the other: GPT-6.1 Sol in Codex Claude Sonnet 5.5 in Claude Code Results (layered app vs. one-file app): Same quality: every visible and hidden test passed in both apps, no regressions Fewer tokens: −21% with GPT-6.1 Sol, −63% with Sonnet ($1.81 instead of $5.34) Faster: −25% with GPT-6.1 Sol, −42% with Sonnet The gap grew with every release: from −11% to −36% (Codex) and from −51% to −77% (Sonnet) fewer tokens Context window: Sonnet needed far more context in the one-file app (162k vs. 96k tokens), Codex didn't Plot twist: before starting a single subagent, Sonnet split the 9,000-line file into 16 modules The resource kit, aka the patterns behind the clean app: Metadata Mapping, Generic Repository, Layer Supertype, Template Method, Factory and a small state machine Honest note: this is one run per agent, app and release. Don't read the exact percentages to the decimal point, but both agents point in the same direction. Chapters 0:00 The claim 0:18 Two apps, same API 0:31 The resource kit and its patterns 1:34 Two releases, two agents 1:56 My hypotheses 2:06 Results: tokens, time, context, parallel work 2:54 Verdict 3:10 Watch them work 3:30 Plot twist and takeaway #ClaudeCode #Codex #AIAgents #CleanCode #SoftwareArchitecture #DesignPatterns