Where are AI coding agents creating the most repo drift for you?

I’m working on a diagnostic suite for AI-assisted software development, and I’m currently looking more closely at where agent drift shows up across different stacks.

One thing I keep seeing is that AI coding agents can complete a task and still leave the repo in a worse state than before. The tests may pass, the feature may work, but the codebase quietly becomes harder to reason about.

That drift can look like bloated files, duplicated helpers, local patches, stale scaffolding, missing verification, or custom code that works around a framework instead of working with it.

The interesting part is that drift does not look the same everywhere. A React project does not drift the same way as a Django app. A Next.js app does not drift the same way as a FastAPI service. A test suite does not drift the same way as infrastructure code.

So I’m curious:

What stacks or frameworks are you using with AI coding agents?

Where do you see the most drift?

What does the agent keep getting wrong?

Are there tools or frameworks where agents tend to bypass the intended architecture?

Are there places where everything looks green, but the repo still feels more chaotic afterward?

I’m especially interested in real examples from people using agents in active repos.