Addy Osmani on AI's '70% Problem' and the Evolving Role of Developers

From the Agentic Engineering Sessions, aired on November 6th, 2025, we hosted Addy Osmani. As part of Google's Chrome Developer Experience team, Addy focuses on AI and developer tools, and he joined us to discuss what he terms the "70% problem" in AI coding.

Over the past two years, Addy has closely monitored AI adoption within Google, where over 30% of code is now AI-generated, and across the broader industry at conferences like Lead Dev. His research has delved into trust metrics, productivity claims, and the practical realities teams face when scaling agentic engineering beyond initial prototypes.

His observations reveal a consistent pattern: AI can rapidly generate up to 70% of a solution, yet the remaining 30% – involving edge cases, security, and seamless production integration – continues to present significant challenges. Concurrently, trust in AI-generated code is declining, even as its adoption rates rise.

For a deeper dive, you can watch the full session on YouTube. Alternatively, read on for key insights and selected quotes from our discussion.

AI Gets You 70% of the Way

Addy explains how AI coding tools are incredibly effective for initial development: "AI coding tools can get you most of the way to a solution, but not all the way. It's deceptively convincing. You can get a UI quickly, write a few prompts, even if you write a PRD you can get something that looks functional. But it can be held together with duct tape behind the scenes." He elaborates on the division of labor: "AI can rapidly produce maybe 70% of the code for an app, for a feature. So the scaffolding, the obvious patterns. But the remaining 30%—things like edge cases where you might have to put in additional debugging, integration with production systems, making sure that your security, your API keys, all of that stuff is in a healthy place—that can be just as time consuming as it ever was."

He also warns against over-reliance: "If you're using AI to generate the code, using AI to test the code, I think that at some point you're probably gonna try throwing AI into the code review loop as well. And at that point, AI is just doing the entire thing. You don't really know what's happening at all."

The Trust Problem

Despite widespread adoption, trust in AI-generated code is surprisingly low and decreasing. Addy notes: "While adoption is in a really good place, trust is surprisingly low and it's declining. Favorable views about AI coding kind of dropped from 70 to 60% within two years, and about 30% of people are reporting little to no trust in AI generated code at all. Which is kind of wild given how much we're relying on this now." He succinctly summarizes the risk: "If you're not paying attention, you can end up—you don't wanna be on Hacker News for the wrong reasons, is a short version of this."

The Two Steps Back Pattern

AI's iterative bug-fixing process can lead to more problems than it solves. Addy describes this: "You'll try to fix this bug. AI suggests a change that seems reasonable. Maybe it's given you a plan in your IDE. The fix is gonna break something else. You're gonna ask AI to fix that issue and it's gonna create two more problems. Rinse, repeat. Sometimes it's five new problems." This cycle can obscure the overall architecture, as he states: "If I ask AI to go and do a thing and it ends up rewriting code in five different places, I have no idea how anything works or how anything connects. A fundamental part of it for me is at least go back and understand how it all works."

Understanding the Code is Key

Addy stresses the imperative of human control and understanding: "The really big difference is that you are keeping the human engineer firmly in control. You are responsible for thinking about architecture, reviewing the code, understanding every line. You have to make sure you're understanding what AI is generating for you, so you are on the hook for making sure the final product is secure, scalable, maintainable." Ultimately, the responsibility lies with the developer: "AI is a tool. If your name is on there, when that code is getting submitted, you are responsible for what you submitted. So you need to make sure that you're reviewing it."

Reality Check on Productivity

Addy provides a reality check on the often-cited high productivity gains from AI: "Often on Twitter, when we see people citing these very high percentage numbers about their productivity gains, if you zoom in, often those are companies that are doing greenfield development on something completely fresh. They don't have technical debt, they don't have all of the baggage that usually comes with traditional software engineering on something that is real and has existed for a while." He estimates more realistic gains: "Developer productivity is 1X, 2X. Maybe they can complete 20% more tasks than they could before. But we're also starting to see side effects. We're actually starting to see that code review is becoming the new bottleneck."

Advice for Junior Developers

For junior developers, Addy offers encouraging advice: "It is in fact really, really motivating to be able to do whatever is gonna help you build something quickly. Make sure you are going back and actually taking the time to understand what it is that was being generated." He champions curiosity: "Be curious. Use AI as a learning aid while still improving your motivation, your feeling about yourself building stuff." He draws a parallel to past learning habits: "Five years ago, we would still see lots of people at every level going to Stack Overflow and in some cases just blindly copying whatever's been most voted as the solution for a problem. Was it actually the right solution? We learn years later, like, oh wait, how did I ship this to production? There are better ways of doing it. So it's okay to be imperfect just as long as you continue to be curious and are open to learning."

Learn More About Agentic Engineering

Explore how to achieve genuine, quality software in an era where code production is no longer the primary constraint. Learn our perspective on Agentic Engineering and catch up on all sessions featuring experts who help develop this crucial topic.

Looking for a Better Editor?

Try Zed today on macOS, Windows, or Linux. Download now!

We Are Hiring!

If you're passionate about the topics we cover on our blog, consider joining our team to help us shape the future of software development.