From PHP to team lead of agents: rethinking judgment, review, and data with Google's Andi Gutmans (Part 1)

In this episode, Eira May and Stack Overflow Director of Platform Engineering Peter O'Connor talk with Andi Gutmans about the throughline between democratizing web development with PHP and democratizing software development with agents today. Andi makes the case that every individual contributor is becoming “a team lead of agents” and walks through how that reshapes code review, interviewing, and the balance of human versus agent judgment. He also explains why he thinks the biggest bottleneck left isn't the models, but the challenge of getting an organization's data into a state where agents can actually reason over it. The discussion also: - Covers how Google is changing its own interview process to evaluate how candidates reason with and guide agents, rather than how well they can hand-code a solution alone. - Explores the "human in the loop, agent in the loop, agent on the loop" framework for deciding where review actually needs to happen - and why that's fundamentally a risk-management question, not a trust-in-AI question. - Uses the Waymo safety-data-versus-perception gap as an analogy for why organizations (and individuals) sometimes resist agent autonomy even when the numbers favor it. - Introduces Google's "borderless lakehouse" concept and the shift from human data stewardship to agent-driven ontology building. Notes: TRANSCRIPT Eira May: Hello, and welcome to Leaders of Code. This is a segment on the Stack Overflow Podcast where we get senior engineering leaders together and talk about the work they're doing, how they go about building their teams, and the biggest challenges they're dealing with right now. My name is Eira May, and I am the B2B editor at Stack Overflow. I'm here with my colleague, Peter O'Connor, who is director of platform engineering here at Stack. Hi, Peter. Peter O'Connor: Hello, how's it going? Eira May: It's going really well. Pretty good for a Monday. Today we have a guest I'm really excited about. We're talking with Andi Gutmans, who is head of Agentic Data Cloud at Google, and Andi is one of the creators behind PHP 3, I think was a college extra credit project, the way I heard it, and now it turns out it's the Bedrock of web development. Andi, welcome to the show. Andi Gutmans: Hey there. Thanks for having me. Eira May: Yeah, thank you so much for joining. I wanted to start with a question for you, Andi, and then I'll let you and Peter kind of take it from here. You helped build tools that made the whole just write code and ship it kind of possible for a generation of web developers. So I'm wondering, how does the shift to agent scale feel from the perspective that you're coming from, the experiences that you've had? Does that feel like a huge change? Does it feel more like a natural progression? Andi Gutmans: To me it feels more like a natural progression. So I think one of the things that was really exciting back in the day when we worked on PHP is we really democratized web development. And actually one of the things that was really impactful is you didn't have to be a computer science graduate to actually use PHP to build a website. So as long as you were a bit technically astute, you could actually build websites. I even had, throughout my career, folks like doctors telling me like, "Oh, I used your language and I built a website for my office," and so on and so forth. And so I think that was a really important time when we made web development accessible to anyone, not only computer science graduates. I think you can think about this moment in a very similar vein where agents are actually enabling lots of practitioners to drive outcomes that are remarkable. My sister, she's not technical, she's a lawyer by degree, but she's been building websites with Lovable as an example, and amazing websites. So you can almost think about this, this is like PHP, but way, way better on pretty much every dimension. And then I think the other piece, the other dimension that really matters is the trust and the security side where I think with agentic development, the best practices can actually be now driven by agents. Where back in my day, one of the issues you had in having non-computer science graduates building websites is they build insecure websites or lots of bad code out there. So I think there's this whole new step up we have right now that is really exciting. Peter O'Connor: Interesting. So I come from the platform engineering point of view, Andi, and it's building tools for those developers. So their lives are even easier. I'm trying to think about when we try to make it easier for them, what are some of the principles when you're thinking about how we think about AI and how to guide it properly and get people to use those tools properly? Do you have certain guardrails or advice that you're thinking about like, "Hey, if you want to do this right, here's some patterns to follow." How do you think about that? Andi Gutmans: Yeah, no, that's a great question. Look, I think the same things matter when you're developing with AI as when you're not developing. There is are you delivering the right level of trust? So are you getting to the right outcomes? It's secure, it's governed. All these things matter just as much, if not more as before. That cost matters a lot, making sure it's easy to use. So I would say that dimensions are probably very, very similar in manner. But the change right now is agents can do so much work autonomously that as we think about the guardrails and how we make sure that we are getting to those right outcomes in the right way, we have to approach that a bit differently. But I think it is still the same kind of fundamental things that actually matter. This is where making sure that you have the right data to activate these agents, that you have the right decision-making around risk-taking on when is it human in the loop? When is it agent in the loop? When is it agent on the loop? You kind of have to make decisions on how you want to operate, but that's actually not that different from engineering decisions that have had to be done in the past on other dimensions. Peter O'Connor: For sure. It's really good to hear the data matters and make sure we guide things properly and the human still matters. A place I hear on my side a lot, which is totally understandable from engineers, is the craft of code is very important to people. And so when I allow an agent to do it, they may not do it exactly as I want because they're not following the pattern exactly as I'd like. Are the patterns in the code craft, in your opinion, just as important as well? Or should we take a step back and say, "Maybe that doesn't matter as much as it used to?" Andi Gutmans: Yeah, I mean even when you look to the pre-agentic development, it mattered less what language you're coding in. I mean, it mattered to a certain degree, but what really mattered was do you understand the business problem? Are you building a sound architecture? Do you have the operational excellence to scale? And so what ended up happening was because you had to do a lot of manual coding, you maybe spent 80% of your time coding 20% on these other things, as opposed to now you can actually spend more of your time making sure you're getting to the right outcomes. So the way I think about this is the value just kind of moves now towards making sure that you're using your judgment on how to guide the agents. You're also making sure that you're reviewing the work that agents are doing in the appropriate manner. For example, I coded something, the agent built a thousand tests for me, and then I had another agent critique those tests and I found out those tests were actually not very good and I had to kind of improve on them. So it still required my judgment. It's just that I was doing less coding as part of that and more designing and orchestrating. And really the way I think about it is every individual contributor now becomes a team lead of agents. Peter O'Connor: Interesting. I think that's a really, really good position to think about things. And I really like the idea of a team lead of agents. How do you see the scope of the knowledge a person needs to have in order to do that work? Because when I think a team lead, you're talking about a senior person. How do we make sure people have those skills? Is there a pipeline problem that we're going to end up running into? Is there a skills problem? Andi Gutmans: I think it's going to resolve itself because I think what's going to happen is even as you think about kids going to college, doing a computer science degree and so on, how we learn computer science is also going to change. Of course, it's critical for folks to have judgment around how a system is built, how a system runs, how to scale. So a lot of these things are going to be the same. But if you think about it, they could potentially take on university projects that are much more complex and a greater scale with agents than if that were hand coded now. So I think what's going to naturally happen is the folks coming out of college are already going to gain that superpower as they go through college, and they're already going to be conditioned to be the team lead of agents. So I definitely think that that's just going to be a natural progression of how things work. It's not that we're not going to need new entrants to the workforce coming out of computer science degrees. Peter O'Connor: Yeah, I really agree with that very heavily. I think I was on a podcast to myself and actually did an interview about interviewing techniques. And I said one of the biggest changes that we're going to have to really look at is it's not so much that you solve some abstract problem very well. I want to hear about reasoning around it and I want to understand how would you orchestrate that problem at a grander scale? And that's hard. Andi Gutmans: Yeah. By the way, we're changing our interview process. So it's not going to be about having an