Switching off your brain

It’s 10am on a Monday, you’re checking Sentry and you come across a crash. It looks simple enough, you only worked on the area of impacted code a few weeks ago and the change you’ll need to make is obvious to you.

The quick fix

Your first instinct is to spin up a new branch, quickly edit the code with the required fix, give it a test run and then submit a PR. You enjoy writing code and nothing feels better than being able to solve something quickly and with very little mental strain, so you don’t even consider giving the task to AI.

Later that same day, you come across a much tougher problem, it might be related to a framework you are not familiar with, the error isn’t clear to you and you go to your LLM of choice, paste the error you’re seeing and within no time, you’re unstuck again and back to writing code.

The entire experience is smooth and satisfying and you’re writing plenty of code, so you must be retaining your skills, right?

The loss of real skills

Until very recently, the above is exactly how I was using these agents, it feels great, you’ve suddenly become a much more capable developer and the days of wasting hours solving obtuse problems is gone.

As time went on I found myself deferring to the agents much more quickly, I knew I could come up with a solution to given problem, but why waste my time doing so when I can ask the agent and continue writing code whilst it solves the problem for me.

But therein lies the problem, you’re never truly thinking anymore. You’re only retaining how to write in the proper syntax, a skill which frankly following the advent of AI isn’t worth anything anymore.

Investing the process

I’m not advocating to be a luddite, AI is undeniably a useful tool and to not use it is not the answer, but we need to consider when we use it.

Taking the original quick bug fix, if you know how the end code is going to look, just give the task to an agent. You’re confident in how the new code should look and will be able to check on it’s work, that said with these latest models, for such a simple issue, you’re very unlikely going to have the LLM produce code you will be uncomfortable committing. Even if you do, you can take this time to practice your code reviewing skills, for such problems you may gain more from reviewing something elses code than just writing your own.

It’s so easy to instantly defer to an agent when you hit a wall, it feels bad to get stuck and you know you can just spin the wheel on an LLM and hope it does the work for you. It’s addictive but you aren’t learning. It’s important to get stuck, to truly think about problems and to come up with the answer on your own. That isn’t to say never use an agent in theses scenarios, they are amazing tools, but maybe let yourself think a little longer before outsourcing your own intelligence.