What is the story about?
The way software is built is undergoing a profound shift. In a recent blog post, Sundar Pichai confirmed that AI is no longer just assisting engineers at Google. It is doing most of the heavy lifting.
Today, 75 per cent of all new code at the company is generated by AI systems and then reviewed and approved by human engineers, a sharp rise from 50 per cent just months ago.
This is not just a productivity upgrade. It marks a structural change in how coding itself is approached. Engineers are increasingly stepping into the role of orchestrators, guiding systems rather than writing every line themselves.
At the heart of this shift is what Pichai describes as “agentic workflows”. Instead of relying on a single tool, engineers now deploy multiple AI agents that can independently execute tasks, collaborate, and iterate on solutions. The result is a kind of digital task force that accelerates development cycles dramatically.
"We’ve been using AI to generate code internally at Google for a while. Today, 75 per cent of all new code at Google is now AI-generated and approved by engineers, up from 50 per cent last fall. We’re now shifting to truly agentic workflows. Our engineers are orchestrating fully autonomous digital task forces, firing off agents and accomplishing incredible things," wrote Pichai.
He shared one recent example highlighting the scale of change. A complex code migration project, carried out by a combination of engineers and AI agents, was completed six times faster than similar efforts just a year ago. The implications are clear. Tasks that once demanded weeks of focused human effort can now be compressed into days.
The company’s experimentation extends to product development as well. The launch of the Gemini app on macOS reportedly relied on an internal agentic platform called Antigravity. Using this system, teams moved from idea to a working native Swift prototype in just a few days. Speed is no longer a competitive advantage. It is becoming the baseline.
Along with it, Google has also announced 2 new AI chips. The TPU 8t is designed specifically for training large-scale models, while the TPU 8i focuses on inference, delivering faster responses for applications such as AI agents.
The company says the training chip delivers 2.8 times the performance of its previous generation at the same cost, while the inference chip offers an 80 per cent improvement.
Beyond Google, the wider tech industry is reckoning with what this means. Sam Altman recently reflected on how dramatically software development has evolved, noting how easy it is to forget the painstaking effort once required to build systems line by line. His remarks carry a sense of both admiration and unease.
Modern AI tools can now generate functional code in seconds, identify bugs almost instantly, and suggest cleaner architectures. Benchmarks such as SWE-Bench indicate that advanced models are increasingly capable of solving real-world GitHub issues with near human-level performance.
What is changing is not just efficiency, but the very nature of entry-level work. Repetitive and structured tasks, once the training ground for junior developers, are now being automated. Smaller teams can build and ship products faster, but the pathway into the profession is becoming less clear.
Coding, in this new paradigm, is less about writing from scratch and more about directing intelligent systems, reviewing outputs, and refining results. The craft is evolving from execution to judgement. And as AI continues to improve, the balance between human intuition and machine capability will define the next era of software development.
Today, 75 per cent of all new code at the company is generated by AI systems and then reviewed and approved by human engineers, a sharp rise from 50 per cent just months ago.
This is not just a productivity upgrade. It marks a structural change in how coding itself is approached. Engineers are increasingly stepping into the role of orchestrators, guiding systems rather than writing every line themselves.
Google AI-generated codes
At the heart of this shift is what Pichai describes as “agentic workflows”. Instead of relying on a single tool, engineers now deploy multiple AI agents that can independently execute tasks, collaborate, and iterate on solutions. The result is a kind of digital task force that accelerates development cycles dramatically.
"We’ve been using AI to generate code internally at Google for a while. Today, 75 per cent of all new code at Google is now AI-generated and approved by engineers, up from 50 per cent last fall. We’re now shifting to truly agentic workflows. Our engineers are orchestrating fully autonomous digital task forces, firing off agents and accomplishing incredible things," wrote Pichai.
He shared one recent example highlighting the scale of change. A complex code migration project, carried out by a combination of engineers and AI agents, was completed six times faster than similar efforts just a year ago. The implications are clear. Tasks that once demanded weeks of focused human effort can now be compressed into days.
The company’s experimentation extends to product development as well. The launch of the Gemini app on macOS reportedly relied on an internal agentic platform called Antigravity. Using this system, teams moved from idea to a working native Swift prototype in just a few days. Speed is no longer a competitive advantage. It is becoming the baseline.
Along with it, Google has also announced 2 new AI chips. The TPU 8t is designed specifically for training large-scale models, while the TPU 8i focuses on inference, delivering faster responses for applications such as AI agents.
The company says the training chip delivers 2.8 times the performance of its previous generation at the same cost, while the inference chip offers an 80 per cent improvement.
AI in coding
Beyond Google, the wider tech industry is reckoning with what this means. Sam Altman recently reflected on how dramatically software development has evolved, noting how easy it is to forget the painstaking effort once required to build systems line by line. His remarks carry a sense of both admiration and unease.
Modern AI tools can now generate functional code in seconds, identify bugs almost instantly, and suggest cleaner architectures. Benchmarks such as SWE-Bench indicate that advanced models are increasingly capable of solving real-world GitHub issues with near human-level performance.
What is changing is not just efficiency, but the very nature of entry-level work. Repetitive and structured tasks, once the training ground for junior developers, are now being automated. Smaller teams can build and ship products faster, but the pathway into the profession is becoming less clear.
Coding, in this new paradigm, is less about writing from scratch and more about directing intelligent systems, reviewing outputs, and refining results. The craft is evolving from execution to judgement. And as AI continues to improve, the balance between human intuition and machine capability will define the next era of software development.
