AI assisted coding requires human expertise
There has been a lot of discussion in our industry about AI assisted and automated coding. Opinions seem very divided, some are swearing by agentic coding tools like Cursor, while others are highly skeptic. Since I work at an AI focused software company I decided to experiment a bit to form an initial opinion.
My experience is that coding agents can help produce code, but it is not without tradeoffs. We are not in a place with human skill (and by extension LLM skill) that we can ship production software without understanding the code. Because of this I find that sometimes speed gains from automatically generating code are offset by the increased time required to review the code.
How to maximise the benefit from using AI coding assistant?
I see AI as extremely helpful as an assistant to get familiar with a large unfamiliar code base, as a precision reference manual tool, or in generating small snippets of code such that reasoning about it doesn’t take too long. I would avoid letting an AI agent autonomously perform multi-file edits in a single prompt.
In my experiments when the agent performed edits over multiple files on a single prompt, I felt that my understanding of the product itself decreased. There was an example of a small bug that was observed in the test environment. While the bug was caught there, its cause was surprisingly difficult to track down, because it was in a part of the generated code that looked entirely reasonable. In cases where I write out the code myself, or when my AI assistant generates the code in small snippets, reasoning about the code’s behavior stays much easier and it is much faster to find the cause of issues.
AI assisted coding should focus on quality over quantity
As a final note it seems like there is a lot of focus on increasing productivity by increasing the speed of which developers produce code. I would like to see more effort channeled into generating correct bug-free software. Achieving this may involve some completely different paradigm than the current LLM trend. I think the future competitive advantage may be an AI that focuses on quality over quantity.
Others have written thoughts about AI assisted coding as well. I found this post to be valuable as it provides concrete tips on how to maximize the benefits of the AI assistants.
Post by Otto
“I’m a software engineer with an interest in artificial intelligence, spanning from machine learning and evolutionary algorithms to logic-driven AI. I enjoy exploring how these technologies intersect with large-scale distributed systems.”