Google Reports 75% of New Code Is AI-Generated

Google says 75% of the company's new code is now generated by AI and subsequently reviewed by human engineers. CEO Sundar Pichai reported that agent-assisted workflows and Gemini models helped complete a complex code migration six times faster than similar work a year earlier. The company is formalizing adoption by pushing staff to use coding assistants and by factoring AI usage goals into performance reviews. Select DeepMind teams have access to Anthropic's Claude Code for specific tasks.

Google developers are using Gemini family models and internal agents to generate, refactor, and migrate code. Engineers retain responsibility for review and integration, but the workflow is shifting toward more autonomous, agentic tasks orchestrating multiple steps. Key operational points:

• •Increased use of agent workflows for multi-step transformations and migrations
• •AI-generated code reviewed by engineers before merging
• •Performance-review incentives tied to AI tool adoption and productivity

This is a major internal scale-up of developer tooling at one of the largest engineering organizations. Adopting Gemini and agentic pipelines at scale reduces developer cycles, accelerates migrations, and creates standard templates for routine work. The combination of models plus agents marks a shift from single-shot generation toward automated pipelines that chain generation, testing, and deployment tasks. Allowing Claude Code for some teams also signals pragmatic multi-vendor strategy rather than lock-in to a single model.

Expect increasing emphasis on toolchain integration, test automation, and guardrails. Review practices must evolve: more focus on security scanning, behavioral tests, and reproducible pipelines to catch subtle bugs introduced by model outputs. Measuring true productivity will require metrics beyond merge frequency, such as defect rates, lead time, and post-deploy incidents.

Monitor how Google implements guardrails and testing for agentic workflows, how performance-review incentives affect engineering behavior, and whether other large engineering orgs replicate the multi-model, agent-first approach. The balance between speed gains and long-term code quality will determine whether this becomes an industry standard or a cautionary example.