I can only speak for my use of it in software development. I work with a large, relatively complex CRUD system so take the following as you will, but we have Claude integrated with MCPs and agent skills and it’s honestly been phenomenal.
Initially we were told to “just use it” (Copilot at the time). We essentially used it as an enhanced google search. It wasn’t great. It never had enough context and as such the logic it produced would not make sense, but it was handy for fixing bugs.
The advent of MCPs and agents skills really bring it to another level. It has far more context. It can pull tickets from Jira, read the requirements, propose a plan and then implement it once you’ve approved it. You can talk it through, ask it to explain some of the decisions it made and alter the plan as it’s implemented. It’s not perfect but what it can achieve when you have MCPs, skills, md files all set up is crazy.
The push for this was from non-tech management who are most definitely driven by hype/FOMO. So much so they actually updated our contracts to include AI use. In our case, it paid off. I think it’s a night and day difference between using base Copilot to ask questions vs using it with context sources.
To the same point as lepinkainen, we are fully responsible for the code we commit. We are expected to understand what we’ve committed as if we wrote it ourselves. We treat it as a speed booster. The fact that Claude does a good job at maintaining the same structure as the rest of the codebase makes it no different than trying to understand changes made by a co-worker.
On your topic of dependency, the same point as above applies. If AI support were to drop tomorrow, we would be slower, the work would get done all the same.
I do agree with you though. I can tell we are getting more relaxed with the changes Claude makes and putting more blind trust in it. I’m curious as to how we will be in a years time.
As a disclaimer, I’m just a developer, I’ve no attachment to my company. This is just my take on the subject.