The Debt the Cleanup Crew Created Every engineering leader knows the deal with a mature codebase. It is a little messy. There are corners no one touches. There are conventions written down nowhere, carried in the heads of the three people who have been there longest. It works. It ships. And when a junior opens a pull request that violates one of those invisible rules, a senior catches it on review and quietly explains why we do not do that here. That was the system. Not the code. The people reading the code. So when the AI coding agents arrived, the promise was obvious. Finally, a way to pay down the technical debt. Write faster, refactor more, clear the backlog. The tool was sold as the cleanup crew. Then someone counted. A large-scale empirical study across 6,275 repositories found that AI-authored code produces 1.7 times more issues per pull request on mature codebases. Logic errors arrive 75 percent more often. Security vulnerabilities show up at 2.74 times the human rate. Input and output bugs at eight times. Here is the part that does not stabilize. The count of surviving AI-introduced issues keeps growing. By February 2026 it crossed 110,000 unresolved. Accumulation outpacing remediation. Not a one-time tax. A curve. The reason is quiet and total. A human writing against that codebase knew which corners not to touch. The agent has none of that. It reads the literal surface and writes against it, confidently, at volume. The friction a senior used to absorb on one pull request a day now arrives on every pull request, all day, with no one carrying the context that made the mess survivable. The debt was never in the code. It was in everything the code did not say, and the people who knew anyway. The cleanup crew was the leak. https://arxiv.org/html/2603.28592v1