Table of Contents
This post is based on my personal experience after dabbling with AI-based agentic programming for a few days. Not that anyone actively reads my blog, but I want to document how I am feeling at this point in time, during what feels like the early rise of AI automation.
How I Use AI Today
I have been consistently using AI chatbots over the past couple of years for programming-related tasks. I barely Google error messages anymore; I just paste them into ChatGPT or Claude and move on. I am on the free tiers now, but I did subscribe for a few months after realizing Claude was better at generating code.
Most of the time, I use these tools to generate snippets so I can avoid revisiting the docs for a library I have used before. I have even dumped LaTeX error logs in whole and just let the model sort them out. There have been times when I dumped entire stack traces into ChatGPT for it to figure out a solution. It has been incredibly useful.
So, Why Not Agentic Programming?
I have never really bought into the hype behind agentic programming, that is, letting an AI model make changes directly in your codebase. The idea that we can trust a non-deterministic machine to make automated edits we would ship to production? That is a big no from me.
Programming is Fun. Reviewing Isn't.
Recently, I decided to move my blog's post path from /articles to
/archive. Since my articles are snapshots in time, I do not plan to
update them unless something is seriously awful. This time, instead
of using my usual Emacs setup, I opened up [Zed](https://zed.dev) and
configured it to use GPT-4 as my preferred model. I asked it to rename
all /articles references to /archive.
It actually did the changes across multiple files. Cool! This was
great, because I did not feel like manually going through
everything. But when I ran my local server, nothing loaded. As I
suspected, the model did not update some related routing logic. I
fixed that manually. Right as I was about to commit, I noticed
something odd: it had changed the word articles to archive in my
about page and even inside blog post content. That made no sense. I
started reviewing everything and, sure enough, found more
inconsistencies.
Now this made me wonder, do I like reviewing code? Ask my colleagues and you would know. I have to force myself to commit the first hour of my day reviewing code to get over the blues. It is not the fun part of my job, and I certainly do not want to spend even more time reviewing changes I did not even write.
Trust, But… Mostly Don't
I will continue using AI to generate snippets, debug errors, and answer technical questions. But a model with write access, editing things in unpredictable ways? That is not my cup of tea, at least not today.