“Everyone I've talked to who uses Greptile, they really loved it, because they felt like they can now trust something it recommends. That became part of review culture.”
VP of Engineering, Datadog
Sesh Nalla
Q&A
At the inflection point of Opus 4.5, I had an all-hands and shared with everyone that I think the output can really go high. The first complaint from everyone was: yes, but what about review? What do we do about it? I don't think we'll have enough time or hands or eyes to go through this.
Our first answer was, let's see who is doing the best here.
Then we were discussing and Greptile came up, and a couple of others. The team was like, yeah, I think we're doing something similar. And I said, no, but let's try. Let's completely give it a try and see what we learn from it. We ran a few other concurrent ones, and we also tried a few artisanal ones — people who put it out there on Hacker News. And Greptile won.
Greptile really picked up. A lot of people started building around it, sharing each other's scores.
I have been watching the channels — I didn't interact with people directly. Some people I asked, hey, how do you like Greptile? And they were all like, yeah, it gets my job done and I don't have to look at all the code. And even in the channels, people were proud of their Greptile score, resharing it. That's how I know.
I actually don't subscribe to people saying they don't want to look at code at all. If a system like Greptile tells you largely everything is great, but these parts you may want to look at, or no, this one can completely go in and you don't have to worry about it — that's still part of your larger correctness suite.
I think people get a lot of that from Greptile. They trust it more.
I would say every part of what you're producing needs to go through a quality process. You want to cut down the amount of bad things you're putting in as early as possible.
There are also social aspects of review. If it is still human-directed, one-person-directed work that needs to be reviewed, then how can the three other people on my team be accountable for the same piece of software? Not only me — because it's the bus factor. If I'm continuously producing stuff and I'm not coming back for a week, being on vacation, and something goes wrong, nobody else wants to take accountability for it. That is also bad.
I think having these validation and review processes (like Greptile) gives others the willingness to take accountability for software produced.
Greptile spread on its own at Datadog
Datadog didn't roll Greptile out. Engineers picked it.
They ran it against a few tools at once, including ones posted to Hacker News, and watched which one engineers kept using.
Engineers started sharing their Greptile scores with each other. That is how the VP of Engineering found out his teams liked it — reading channels, not running a survey.
Review is what makes someone willing to be accountable for code they did not write. Sesh Nalla's point is that this does not change when the code is AI assisted.
Learn more about Greptile
Greptile consistently catches more bugs in large repos
