Google just crossed a threshold that should terrify and excite every software company: three-quarters of its new code is now written by AI.
The Summary
- 75% of Google's new code is now AI-generated, reviewed by human engineers, up from 50% last fall and 25% in October 2024
- A complex code migration using AI agents alongside engineers finished six times faster than engineers working alone could manage a year ago
- Engineers now have AI usage metrics factored into performance reviews, and the company is pushing "truly agentic workflows" where autonomous tasks run without constant human supervision
- Some Google DeepMind employees have been using Anthropic's Claude instead of internal tools, creating internal tension
The Signal
The velocity curve here is the story. Google went from 25% AI-generated code to 75% in roughly six months. That's not incremental adoption. That's a phase change in how software gets built at one of the world's largest tech companies.
CEO Sundar Pichai framed this as a shift to "truly agentic workflows", which is corporate speak for "our engineers now manage AI agents that write code instead of writing most of it themselves." The 6x speed improvement on that code migration isn't just about typing faster. It's about parallelizing work that used to require sequential human attention.
"A complex code migration done by agents and engineers working together was completed six times faster than was possible a year ago with engineers alone."
Here's what's not being said loudly: Google is restructuring engineering work around AI tooling whether engineers like it or not. Performance reviews now include AI adoption metrics. You're not just encouraged to use Gemini for code generation. You're measured on it. That's a forcing function that will propagate through every tech company within 18 months.
The internal friction around tool choice matters more than it seems. Some DeepMind engineers preferring Claude over internal Gemini tools suggests even Google's AI researchers don't think Google's AI is always the best choice for their work. When your own elite team is shopping elsewhere, you have a product problem or a trust problem. Probably both.
This also reframes what "engineer" means. If 75% of the code is machine-generated, the human role is shifting from author to editor, from builder to reviewer. That's a different skill set. Pattern recognition over syntax memorization. System design over implementation details. Knowing what good code looks like matters more than knowing how to write it from scratch.
The Implication
Every engineering org is about to face the same pressure Google just normalized: why aren't we moving this fast? The 6x migration speed becomes the new benchmark. Boards will ask CTOs why their teams aren't using AI agents this aggressively. Investors will compare your velocity to companies that have.
For engineers: your value is no longer in writing boilerplate or even complex implementations. It's in judgment, architecture, and knowing which problems the AI is solving wrong before the code ships. If you're not learning to work with AI agents now, you're learning a job that won't exist in two years. Watch what tools the best engineers are using, not what companies say they should use. The Claude vs Gemini tension tells you everything about where the real edge is.