Working with AI Agents - Know how to challenge them

As I have shared in some of my previous articles, I have been working a lot with the AI agents. I have to say that instead of me doing any kind of code, my day-to-day work is more like reviewing the code created by AI Agents. I have now transitioned to this. Even before this, it is a regular thing that I review a lot of code written by my team members and approve it or improve it. But in such cases, I also sit and write some code by myself. But today the world is moving toward a future where no human will have to sit and write lines of code. It can be handled by the AI, and the only requirement is strong supervision to ensure the code written by the AI is intact and correct.

This has mostly become my job. I wouldn't say that it is an easy job to do. Even though AI writes the code for me, I have to sit and review that code line by line and challenge it. This has become my routine, and that is also because AI is stupid in many things, and it can make a lot of mistakes. We can correct those mistakes only by giving a prompt back, challenging the AI. In order to do that, we have to be very strong in logic and reasoning. We cannot accept any code that is written by the AI. We have to challenge the AI so that we make sure we get the right output.

Source - Sora

In the last few weeks, I have seen multiple instances where AI makes stupid mistakes. All because of some assumptions. But with the latest models I have seen, their Hallucination problem is somewhat contained, and inside the same context, they can deliver quality code without any evident defects. The problem is mostly with the approach they take to solve the problem. The following is just an example of what I'm trying to convey. If we need 5 times 2. We have to use multiplication because it is the most efficient way to get the answer. But sometimes, if 2 is added 5 times or 5 is added 2 times, that's not the efficient way, and I have seen AI making such mistakes, and if we ask AI back Why can't you do multiplication, then it replies, understanding that this solution can be arrived at with multiplication as well. I'm not saying this as the exact problem because this multiplication or addition decision-making is something AI can do a lot better. I just took it only as an example to explain my thoughts.

Sometimes it is also draining to work with AI agents because they give us solutions and code that is far more advanced. The approach that it takes is also very high quality, and in order to validate the code, it can be exhausting because we have to first understand that particular concept to review the code. So there is a lot of learning in working this way, but only for people who are already good at programming and those who review the code completely generated by AI and challenge the agents with further questions.

If like what I'm doing on Hive, you can vote me as a witness with the links below.