Google Relies on AI for 75% of Its Code, Yet Engineers Now Deliver ‘More’

The emergence of agentic systems has automated much of the day-to-day work developers handle. Hence, the software industry is confronting an uncomfortable question: If AI can write code, what happens to developers? 

At Google, the answer is not fewer engineers, but giving more output. 

More than 75% of the company’s code is now generated by AI, according to Richard Seroter, Senior Director and Chief Evangelist at Google Cloud. But instead of shrinking engineering teams, Google says developers are producing more software than before.

For developers, this shift is already changing the nature of the jobs, while companies are expecting engineers to think more about systems, outcomes and product decisions. 

"The smartest companies are not waiting for this to settle," Seroter tells AIM. "They are saying we have to be in the game, we have to be rethinking our work, knowing that we're going to throw some things away. But if we wait, we are going to lose out." 

The Outcome-Driven Era

At the heart of Seroter's argument is a shift in how developer success is defined. The old yardsticks, the elegant algorithms and manual coding ability are becoming less important than the ability to deliver outcomes quickly. 

AI tools are accelerating this reorientation by design. "When you prompt Gemini or Claude or GPT, you prompt with an outcome," he says. "We are training ourselves to think in outcomes, and that might be better than training ourselves to think in lowest-level algorithms and syntax."

Google measures its engineering teams against three internal markers: speed, ease, and quality. AI, Seroter argues, should improve all three simultaneously. 

Seroter notes that this did not mean engineers were doing less work, but more, including experiments, architecture, and systems thinking. Some were doing tasks that previously belonged to product management or user experience. He says Google's goal was not to reduce staff but to triple their output with the same team.

Simpler to Build, Harder to Operate

One of the more nuanced points Seroter raises is that AI simultaneously simplifies and complicates the developer's world. On the simplification side, he is enthusiastic and says that barriers that once killed projects, such as a missing answer on the internet or an unfamiliar framework, have largely disappeared.

"I have never built front ends in my life," he admits. "Now AI just makes it easier to build exactly what I want. He says there were many times five years ago when developers would have given up because they hit a stumbling block, but that doesn't happen anymore.

But the ease of building complex, multi-agent systems has outpaced the ease of running them. "The work to build is easy, but the work to operate is hard," Seroter acknowledges. "Unless platforms get better at troubleshooting, we are just moving the work elsewhere." 

He points to Google Cloud's efforts to surface errors intelligently, pinpointing the exact line of code that failed, or helping developers navigate sprawling architectures, as evidence that both sides of the equation must be solved.

The Prerequisites No One Wants to Do

Perhaps the most overlooked aspect of Seroter's argument is the essential, yet unglamorous, foundation that effective AI workflows rely on. He believes that data management is a sector of enterprise that is not investing enough in agentic systems. 

"If I don't have good data management and I don't trust an agent to access a database and pull data out, this is not going to work," Seroter adds. He says that if he hadn't considered giving clear instructions to guide the models and had believed he could simply purchase AI and begin building agents, he would have overestimated the results.

API management falls into the same category. Seroter says that most companies are hoping to skip these prerequisites, but this is one trend where shortcuts will not hold. 

"This is the first trend in a long time where you cannot cheat," he notes. "The AI results won't feel as good because they are not passing in the right stuff. Folks will realise that."

Skills, Judgement, and What AI Cannot Replace

Despite the enthusiasm for AI-assisted development, Google is firm that developers cannot outsource their judgement to models. The ability to read, evaluate and understand code, even without writing every line, remains essential. 

"If you don't have the experience, you don't get the judgement," Seroter says. He says that despite not having written a line of code since December, he had never produced more software. He worried that they (the developers) might lose the ability to understand what is happening if prompting became the sole skill.

Seroter describes the ideal developer as a T-shaped professional, combining deep expertise in a specific area with a broad range of skills. For those just starting their careers, he suggests focusing on relevance and versatility rather than narrow specialisation. 

The ability to use AI tools to build remarkable things is worth more than mastery of a single language. "It is a much tougher market than it was five years ago, where every engineer found a job immediately," he says and calls on universities to rethink their curricula accordingly.

As AI reshapes software development, coding may become easier, but building reliable systems, managing data and exercising technical judgment are becoming more valuable than ever.