A story about how Claude, an AI coding agent, solved a complex Terraform schema bug in mirrord not by brute force, but by reasoning its way out with a helper script it wrote for itself.

MetalBear

An engineer shares their journey from skepticism to trust in AI coding agents after Claude solved a complex Terraform schema validation bug. Unlike previous AI tools that looped endlessly, Claude broke through by writing its own helper script to optimize its debugging feedback loop, demonstrating problem-solving capabilities beyond simple code generation. The breakthrough came when Claude analyzed the schema for specific error patterns rather than relying solely on the provided test case, showing reasoning ability that resembles human engineering approaches.

How I Learned to Stop Worrying and Trust AI Coding Agents

As an AI dev i have to say:
Please worry and please dont trust ai coding agents!
A friend of mine just deleted a prod db because he trusted ai coding agents. Thankfully he was able to restore it from some random temp files. (This guy is not a dev and the project was a self hosted ticketing software)
AI can do amazing things but it needs strong human in the loop functionalities and preferably also functioning backups.

I don’t even trust myself to code 100% accurately. Every line is a potential liability, and the best software development is an iterative process that requires constant micro-decisions all along the way, whether I am the one typing the code or an LLM is generating it.
Yes, you will have some mind blowing successes with these tools; they have access to and can parse through more information than you’d ever keep in your head. Just by that alone, statistically, they should be able to come up with some novel solutions that you or I have never seen before; that’s the main upside to the “generative” part. You will also encounter some of the most disruptive, broken and terrible outcomes that will waste more time than had you just put in the work and done it yourself. Such is the case with non-deterministic outputs, right?
If you’re “trusting” anything, even your own mind, you’re missing the most key component of this discipline: relentless inquiry. Even when I come up with what I think is the absolute best solution, I am still left wondering: could it better? What could I be missing?
To skip this part and instead “trust” anything, especially an algorithm, you should probably put the keyboard down and do some soul searching whether you think you still enjoy this work in the first place.

Then your code can quickly turn into spaghetti 😂

I don’t know. It could also be down to the inherent randomness of LLMs or getting familiar with how to use the tools. I started using Cursor (chat mode) exclusively about 5mo ago and experienced similar problems in the beginning. As I became more aware of the tools capabilities and limitations, the problems went away more or less completely. And I experienced it also writing test scripts to aid its debug loop. Chat is probably not as slick as an agent though.

I read the article title and rolled my eyes. There is no way a developer should trust the output of an AI suggestion. AI has to be directed and reviewed. What crappy advice! But then I read the article and found this more reasonable quote:
“I still don’t think coding agents are ready to run wild on a codebase unsupervised. They need direction, they make mistakes, and sometimes they fail in ways no human would. But after this experience, it’s clear to me that AI coding agents are starting to move beyond autocomplete.”
The title was clickbait. The article was fine