Over time the codebase will get more and more bloated, and making changes will take longer and cost more tokens. You'll start losing your grasp on how the codebase works, to the point where you can't even tell if the agent is doing the right thing or not. You'll be lulled into thinking that the code is correct from the tens of thousands of AI generated unit tests and the 100% line coverage, then you'll start noticing bugs pop up when you actually try to use the product. You tell the agent to fix the bug. It makes a patch and adds another dozen regression tests. You re-test the product. The bug has been fixed, but now there's a new bug. This is your life now.

And then you reach critical mass (http://www.catb.org/jargon/html/C/critical-mass.html), when trying to fix each bug only increases the number of bugs.

And Anthropic will happy sell you the newest model, tuned specifically for these problems. At a premium of course.