I saw this image posted on X with reference to a development team losing substantial volumes of work due to agent hallucination.

I noticed something similar in my work, and I really emphasize the similar here.

It’s similar in the sense that there is complexity of all sorts being added continuously to codebases. Some of that complexity we can immediately notice, some other, we don’t.

Here is my comment on X to that image:

I understand the problem with not knowing what a function technically does. I believe it can happen and may be a big problem.

I faced a new (big) problem this past week that is similar and equally concerning. I’m writing this because I can connect it somehow to the problem described in the image.

I’m calling it “functional misalignment”.

The agent was working on a Controller and then realized it lacked a set of sub-workflows and data that were needed for the main function to return true. They were missing simply because the system itself is incomplete… I am building it.

So the agent decided by itself to build those resources on the fly to meet my requirements. And here comes the “misalignment”: it created arbitrary business assumptions that didn’t exist and the result was a (business) functional misalignment.

I noticed the side effects a bit later when the agent started to ask me very weird questions within the context of other tasks. I simply could not understand those questions. The agent was attempting to build more misalignment.

I was lucky to catch it in time and fix it.