This is part of a series of blog posts that stemmed from my first experiences with “vibe coding”. For more on what I did and where you can find the rest of the posts, please refer to the first post.
While AI Assisted Coding is really powerful, one big limitation is that it can only help you with what it can see. Specifically, it’s trapped in your IDE (or at least within your laptop). I found this troubling in several different areas:
- Testing and debugging – Once my code changes were complete I spent the next little while pushing changes and then running my pipeline and copying a set of compile errors defects errors back into Claude for help debugging them. The whole process would have been much more seamless if Claude had direct access to my CI/CD pipeline.
- Setting up infrastructure – LTHOI runs on Lambda, DynamoDB, API Gateway, and Queues. These are all setup in a variety of CloudFormation Stacks that are defined in the git repo. While Claude can see those, it cannot see anything in the Amazon environment. This is a major setback that AWS needs to address ASAP. Even Amazon Q can’t see configured AWS stacks. This is too bad, I had a small problem with one of my stacks and it took me a while to find it. I feel confident that even a pretty simple AI assistant running within my AWS console could have helped. Even better if Claude could have been given access.
- Finally, AI assisted coding is not a solution for SRE. It’s only being used while I’m coding. There are options for GenAI in the runtime environment, but they are not as directly linked to the software development as I would eventually like to see.
There are solutions to this problem of course, the most common of them being the development of MCP servers and agents (I’ll discuss this in more detail in a future post). It’s important for enterprises to remember, even if there are ways around it, they will take time, effort, and skill to implement. Just like with DevOps, test automation, IaC, and all the other technologies that have made software development more effective this one has not (yet at least) made it effortless.