AI Now Produces 75% Of Fresh Code at Google, Confirms Sundar Pichai

Software development is undergoing a significant change. Sundar Pichai affirmed in a recent blog post that AI is no longer limited to helping Google engineers. The majority of the hard lifting is being done by it.

AI systems now produce 75% of the company’s new code, which is subsequently examined and approved by human engineers. This represents a significant increase from 50% just a few months ago.

This is more than just an increase in production. It signifies a fundamental shift in the way that coding is seen. Instead of writing every line individually, engineers are increasingly taking on the role of orchestrators, directing systems.

Related: Indian AI Stocks You Should Watch Out for in 2026

What Pichai refers to as “agentic workflows” are at the center of this change. Engineers now use several AI agents that can independently carry out tasks, work together, and refine solutions rather than depending on a single tool. The end product is a type of digital task force that significantly speeds up development cycles.

“At Google, we’ve been utilizing AI to create code internally for some time. At Google, engineers now approve 75% of all new code, up from 50% in the fall. Now, we’re moving toward genuinely agentic workflows. “Our engineers are firing off agents, coordinating fully autonomous digital task forces, and achieving amazing things,” Pichai wrote.

He gave a recent example that demonstrated the extent of the transformation. Using a combination of engineers and AI agents, a complicated code migration project was finished six times quicker than comparable projects only a year ago. The ramifications are obvious. Work that once required weeks of concentrated human labor can now be completed in a matter of days.

The corporation also experiments when developing new products. According to reports, the Gemini app’s macOS debut depended on an internal agentic platform known as Antigravity. Teams used this technology to go from an idea to a functional native Swift prototype in a matter of days. There is no longer a competitive advantage to speed. It’s becoming become the standard.

Google has also revealed two new AI chips. While the TPU 8i concentrates on inference and provides quicker responses for applications like AI agents, the TPU 8t is specifically built for training large-scale models.

According to the business, the inference chip offers an 80% improvement, while the training chip gives 2.8 times the performance of its prior version at the same cost.

The whole tech sector, not just Google, is considering the implications of this. In a recent reflection on the significant evolution of software development, Sam Altman noted how easy it is to forget the meticulous work that was once necessary to design systems line by line. His comments are both admiring and unsettling.

These days, useful code may be produced in a matter of seconds, defects can be found nearly instantaneously, and cleaner structures can be suggested by AI technologies. SWE-Bench and other benchmarks show that sophisticated models are getting closer to human-level speed when it comes to addressing real-world GitHub problems.

Not only is efficiency changing, but entry-level job itself is also changing. Once used as a training ground for young developers, repetitive and regimented chores are now being automated. Although smaller teams may produce and dispatch goods more quickly, the career path is becoming less obvious.

In this new paradigm, coding is more about guiding intelligent systems, analyzing outputs, and improving outcomes than it is about creating from scratch. From execution to judgment, the art is changing. The next phase of software development will be defined by the harmony between machine intelligence and human intuition as AI advances.