What Is "Agent Debt" and Why Is It Breaking Your AI Code?

Since our last episode, you've enjoyed a front row seat to just how quickly software development has changed over the past year alone thanks to AI.

So let's start there.

We began by, uh, using AI to help us write code, right? But now many teams can generate software faster than they can realistically review it.

So are you seeing the bottleneck shift from writing code to a bit more downstream into the reviewing, deploying, and operating it? And also, is that shift changing the role of software engineers and SREs?

Yeah, what a wild time it is to be in the software space.

I mean, so one of the fun parts about being at New Relic is that we're not just building software tools, we're of course a software shop ourselves.

And so, you know, we're customer zero for everything we build, and a lot of what we've ended up kind of releasing ultimately comes out of things that we face day to day.

And so we do have that front row seat, you know, and I've been able to work with the teams through this journey.

A year ago, like you said, it was mostly very fancy auto-complete.

There was some grumbling.

People are like, "Oh, I'm tabbing my way through software." "Oh, I'm experimenting with an agent here or there." Recently, I talked to some of our top engineers, and we asked them, like, how they write code, and I still remember this woman, she said, she said, "I haven't written a single line of code in calendar '26." Like it's been over six months since I opened my editor for typing code into.

And we've just completely changed that whole system from having kind of this extension of your fingertips it... that we had in the early era to having all of these agents running, sometimes multiple, like, agents simultaneously, running agent teams, running, like, loops endlessly working on the software.

And so we've just poured all of this acceleration onto the ability to create more software.

But when I switch to my operations hat, more software is not what I want.

More software is the problem.

That's not the solution, is that, you know, you've got now this huge pile coming at you as an ops person, and if something doesn't do what you think it should, you often don't even really have somebody to ask.

And, like, it's this just, like, completely new universe v-versus where we were 12 months ago.

I mean, seven months ago.

I mean, it's...

Everything changes so rapidly now, and I think that what we've discovered and what e-everybody needs to be doing right now is thinking about what are the ways that the software development process has to change, and the way we think about AI, you know, tools in particular needs to change to move out of that creating endless piles of software and into a better system to ship features.

Yeah.

I think you talking about this, and this goes on another conversation we can have for another day, but I also worry about, like, the... how software engineers, how anybody working with code can stay on top of their game when you're just going into this, like, auto-complete, you know, zone.

That's an entirely different challenge that I think many of us are gonna [chuckles] have to figure out how we can balance it without over-relying it, but... over-relying on that.

But one thing that everything that you're talking about, this actually ties into some research that you've recently released from New Relic.

You found that compared to pre-AI adoption, 94% of leaders rated AI-generated code as higher quality at review time, but 78% report more incidents once it ships Why is this happening? Where is this discrepancy coming from? And, and also linked to this, is this where this phrase of agent debt comes in? What's the first sign that agent debt is building up on a team?