Who reviews your agent's code? | Santosh Yadav, CodeRabbit

This talk addresses the evolution of code reviews in the era of artificial intelligence and the limitations of traditional GUI tools. The speaker discusses how the introduction of AI agents has accelerated coding processes, often outpacing human reviews and creating bottlenecks. They explore different code review methods, including command-line interfaces and integrated development environments, emphasizing the community's shift towards more effective solutions. The speaker introduces Code Rabbit, a tool designed to enhance code review by organizing files logically rather than alphabetically, thus improving collaboration among engineers. With features like contextual comments and an integrated AI agent, Code Rabbit aims to streamline the review process, especially for open source projects.

So, I'm going to give you idea about like how people are doing a review. For to start with like it's been 16 years. Who remembers that? Right? But it was 16 years ago, not 16 18 years ago when GitHub was launched. Right? And then we have been doing the review the same way, the GUI. I I'm not sure like how people are doing it, but that's how

we did it for almost 16 years till 2024. And then agents change how we write code. It changed for everyone, right? Everyone is now more productive, more fast. Let's I mean yeah, we can we we can agree to disagree that we are more productive, but your company never agrees to that. That's why we are not getting raised. And then agents are actually writing code faster than someone

can review and it's now a bottle neck bottle neck, but people are from the community, they're actually trying to revisit how we should be doing the code review. Let me show you a small UI. Ah, okay. believe me this UI is not for humans. Forget about agents. No one can read this. But this is how we are doing code review right now. Most of us are actually

using an AI review tool and or even in-person review or someone your senior who is doing a review. And you have so many comments. No one can read it. So, how community is actually catching up with the change of the era which is AI era now. So, there are multiple ways people are doing it. Of course, people love CLI, right? This is where it started some months

ago. I mean I just It's moving so fast like it's just it was just few months ago people were saying, "Oh, you know, IDs are dead." And everyone loves CLI. So, people were doing CLI. So, people were running the code review in the CLI. So this is still there are still people who are doing this. So this is one way to do your code review, right? Just

open the CLI and run your coding agent or review agent and then it does the review. And the advantage is you don't have to actually go to that nice view I just showed you few moments ago and read through the entire line of code which sometimes just agree let's let's agree most of the times doesn't work, right? What next? People are still there who loves IDE. And

now people are actually also doing or using the IDEs to go do the code review. So this is another set of people who are like okay web doesn't work anymore for larger diffs or larger PRs. So let's do it using IDE. The people who are still doing it using code X cloud code, right? They can give you the show show you the diff and you can do

a your review there and then maybe talk to your friends like what needs to be changed. And now agentic era, right? So people are like oh look I'm not even looking at my code. So I'm all cool. And now all the code review agents also supports the agentic way of writing or reviewing the code where they can actually just run in the loop. You can just run

the agent it does the review pushes the fix checks again until unless there is a zero review comments and then you are all happy because your UI look really nice because there are no comments to solve. So this is another way. But we still think there needs to be something which is for both human and So at Code Rabbit we are actually trying to change this view.

The same view which you saw in GitHub looks looks more slick and cleaner with the new view new view actually we are trying to do. So here, what we are doing is we are trying to bring all the context about the code changes, all the files you can see, but for ages GitHub has been giving us the same view where it's in alphabetical order. The files are

alphabetical order. I'm sure like you might have all noticed that, but we are trying to actually bring and club the files together as an engineer might have written that. So, it starts from the files which probably I as an engineer started creating first. For example, maybe models, right? And then I went ahead and wrote services, and we try to do the same here by using something the

concept of something known as layers. So, if you see there on the right-hand corner, it's called layers. And the left-hand side, we are trying to show you all the comments, all the context, all the informat- information about the code. And we also have our AI agent available there, so you can talk to agent and see like if something needs to be improved or something needs to be

maybe not fixed, and you can do all in a single view. So, this is the new view, new view we are trying, and Code Rabbit is free for open source, so you can actually just go ahead and give it a try today. You don't have to even come to me and pay my pay money anything. So, just go ahead and try. It's free for open source, and

if you are an open source project, just talk to me. Uh maybe I can do something more for you uh because we have $1 million fund to support open source projects. And we have been already doing that for lot lot so many projects. But, there are still old-school engineers we uh who has been doing this. I'm sure most of us do this. Looks good to me, and

let's go. Yeah? And then there is a staff or principal engineer who looks at your single-line PR and gives you 50 comments. Then you are like, "Why?" But, yeah, people exist, right? And And there are web coders who are actually Oh, who needs a code review? My agent is actually doing perfect job. It writes perfect code. And after few weeks this they think about where is my

database gone? And yeah, the cycle continues. And if you have time you can scan this code and tell me like how you are doing the review, but not required. You You can just skip that, but I'm also done with my talk.