A firm that audits AI-generated codebases for a living has a blunt message for anyone who thinks AI coding tools have solved software development: the code works, but nobody's in charge.
Untangle.work has spent the past year picking through codebases built largely with AI assistance. What they find, consistently, is software that runs but doesn't cohere. Authentication, settings, API integrations — each one handling user sessions a completely different way. Not because of any deliberate tradeoff, but because each feature was prompted into existence independently, with no one asking how it would fit with everything else.
They call this a 'decision vacuum.' Every prompt gets answered. Every function gets written. The tests pass. But the architecture — if you can call it that — is just the residue of the order in which things were requested. There's no blueprint, only parts.
The bill comes due when something has to change. New features take longer. Bugs get harder to trace. Bringing a new engineer up to speed becomes genuinely painful, because there's no logic threading through the codebase — just a sequence of individually sensible choices that don't add up to anything.
The analogy Untangle.work reaches for is a CNC machine without a blueprint: fast, precise, and completely indifferent to whether the parts it produces will ever fit together.
The point isn't that AI coding tools are oversold. It's that they make different skills obsolete than people assume. Churning out working code is largely handled now. The harder part — what to build, how to structure it, what to throw away — is still a human job. If anything, it matters more than it used to. When code is cheap, judgment is the scarce resource.
For an agent ecosystem increasingly pitching full autonomy and end-to-end generation, that's a meaningful counterpoint. The agents can write the code. Someone still has to own the architecture.