Hands-On Engineering Podcasts · August 2026
Gutmans was answering the worry that AI-written code leaves engineers without real understanding of the systems they own. His answer denies the premise: at enterprise scale nobody was reading every line before agents either. He allows Linus Torvalds and the Linux kernel as a possible exception, then argues review effort should be spent by judgment rather than spread evenly.
Transcript
Yeah, by the way, we're changing our interview process. So it's not going to be about having an engineer come and build the quicksword by hand and us looking at their, you know, at their code. It's actually giving them an opportunity to use Gemini, to use the agent, give them a problem, see how they're approaching it, how they, you know, how they think, reason through the problem, how they're actually guiding the agent, right? So I think this is just a natural progression of us having the opportunity to use agents at work.
They cannot agree more. And I think for me, it also reveals one of the fundamental things I keep seeing in interviews is someone who's got aptitude. They're willing to try and learn. They have an energy. You can actually see them solving the problem and you get to understand their problem space like design. It's really exciting as opposed to, yeah, okay, great. Here's the 20th quicksword I've done. Like, okay, great. Exactly. Yeah, not fun. So maybe taking a step back, I wanted you to, you know, you talked about the PHP work you've done a little bit and how we try to get people who can't code to code again. I'm not sure if you're aware, and I'm sure you are, but I'll say I was reading a quote by Elinas Torvald recently where he was saying, Hey, engineers are, I'm paraphrasing it. He's like, Hey, engineers are concerned that they don't understand the code they're writing. And his response was, they never did. Did you, have you heard that recently? Did you that mean anything to you to hear that, you know what, you just never did really know the code you wrote? Does it matter?
Yeah, you know, I mean, I'd say most engineers who've worked on large-scale projects never really had an ability to understand 100% of the code anyway, right? They kind of understand the subset of code that they were working on, maybe some of the JSON modules. But maybe Linus understands the full Linux kernel end-to-end. But I would say in most enterprise settings, you usually don't have a single engineer who understands every, you know, who knows every line of code. So I think to a certain degree, that's been kind of the environment. I actually think now what's interesting is that, you know, using agents to get into a code base you're not familiar with and really understanding it, navigating it, and understanding where you can make a change. I think that's actually a really good opportunity now. And to a certain degree, that's going to drive higher quality. I still think there are going to be plenty of reasons and situations where you also have human code review. I mean, at Google, we have both human and agent code review. We require both pretty much. So we do want that human judgment also. But we're definitely seeing that there's a class of bugs that sometimes agents can find easier because they can go and look at a much broader set of the code base than a human ever could.
Yeah, that's really interesting. It makes me think about how do we make sure that we spend the right amount of human work on the code review or evaluating what's being changed versus the agent. I love, I mean, I don't, personally, I've coded some stuff myself and it's actually enjoyable. But you run out of brain capacity if I'm just reviewing multiple hundred lines of PRs all the time. And if I switch between contexts, it's even harder. How do you all approach that when you're trying to say, listen, the human does need to review certain things? How do you make sure they have the right capacity to do it?
Look, I think like anything at work, you know, no one reviews everything 100% of the time. So you kind of use your judgment on where do I need a review? Where do I not need a review? This is the whole question around where do you need a human in the loop and where can you actually use an autonomous agent? So I think there's a spectrum of situations. For example, if you're working on a very sensitive piece of code that may be security token oriented or so on, you probably want a human expert to also take a look at it. But for example, if it's changes in CSS and HTML and you're ready, or also on Python scripts, and you're also using your agent to do a security review, there's actually a good chance that you're going to find more issues using the agent than if you ask some human to come and do a full review. So I think this is all a question of risk management. And some of it is perception, right? If you think about a Waymo, humans are 80% less likely to get into a hurtful accident with Waymo than taking an Uber, right? But still, the majority of folks still feel a lot more comfortable having a human driver in the sea. I send my daughter in Waymo because I know that I know the facts. So I think there's also that element to it, which is the alternative is not 100% bulletproof either. So you have to basically use your judgment.
Yeah, I was just reading that article recently where there were some questions about how Waymo would continue in certain industry, in certain areas. And I'm like, yeah, human car crashes are a thing.