Ever since LLMs were introduced to the mass market, they have started to revolutionize the programming space. It was marketed primarily as a replacement for software developers. I still remember the time when I opened ChatGPT for the first time and was carried away with awe. I started to wonder what all the possible things were that could be done using LLMs.
AI agents
AI is just a tool and a tool cannot be held accountable!
By 2026, things had improved a lot in every way. Now we have AI agents that can operate like humans, access our file systems, and perform tasks with higher accuracy. Developers are now expected to deliver outcomes very quickly by companies, provided that they have OpenAI or Anthropic premium memberships. Even though the feat of AI is remarkable and its ability in code generation has been skyrocketing, the quality of the code remains a big question in most companies. People still don’t understand that AI is just a tool, and a tool cannot be held accountable for a given project. So vibe coding is not the problem, your code review is.
Need to get basics stronger
One major problem I see with software engineers who are in the early years of their career is that they are not strong enough with their basics. I don’t know how some companies trust code generated through an LLM by a fresher graduate within a week, when it usually takes 3 months for a seasoned developer to produce the same. But the actual issue comes only after that: the system breaks and a production fire at midnight cannot be handled by the developer, as he has little knowledge of the code that has been deployed. The solution here is to get strong on the basics of computer science first, and also to have a watchdog system that does not let the developer push bad code to production.
Manual Code Review
Reviewing code, whether written by yourself or by a fellow peer, is a primary responsibility of developers. I always tell my juniors to do a self-review before raising a pull request. This saves a lot of time for the reviewer and also speeds up the push to production. Even if it is AI-generated code, I would still lean toward reviewing the code myself really carefully before committing. The same responsibility applies to the peer reviewer as well. He might have a lot of tickets to complete for himself, but he must take the time to review the PR carefully and ask the author to make the required changes. Now, the challenge here is that a single PR nowadays can easily contain more than 2000 LOC, which is where AI review comes into the picture.
AI Code Review
AI code needs an AI reviewer too
No human can actually do a thorough code review for a 2k-3k LOC PR. That is simply impractical. Every engineering team in a company should be setting up a detailed AI workflow to handle this issue. An AI workflow is simply a group of AI agents present at each stage of the SDLC to refine the work of the other AI agents. This reduces the overhead for humans to verify each task. For reviewing code, one can set up a reviewing agent directly on their own version control software, like integrating a Claude agent into GitHub. This setup saves developers time and also improves code review quality. I will share my AI workflow that use everyday in my job in the upcoming posts. Stay tuned!