At work, I use Claude, ChatGPT and Copilot, depending on the task I have. Sometimes I create my task plan using one agent, another one helps with writing the code and another one does the code review. Sometimes I use the same agent for all three steps, so I don't really have a flow that's set in stone.
One thing I did notice though is that, all agents perform very well when it comes to generatig code for greenfield projects but the quality of their output decreases the more I work with legacy code. This is something that my colleagues also noticed, so I assume my experince is not unique.
Since our code base is 100% legacy, I do use the agents for generating code, but what I use them for the most is explaining code. And that saves me a lot of time and frustration. In the past, if I was lucky enough and the person that wrote a very complicated piece of code would still be working in the company, I would ask them to help me understand their reasoning.
Now I do not really need to do that. If the documentation is not sufficient, I can always ask an agent to explain the logic behind some code and almost everytime it comes back with a reasonable explanation. Sometimes the explanation is really silly or it's a straight up hallucination but I think that's where our common sense should kick in: while reading something the agent outputed, by looking at the code we should be able to tell if the explanation makes sense or not.
