Google: 75% of New Code Is AI-Generated, Code Migration 6x Faster with Agents
Google announced that 75% of new code written inside the company is now AI-generated and reviewed by human engineers. The number has steadily increased from ~25% in October 2024 to 50% last fall, according to a blog post by CEO Sundar Pichai on April 22, 2026.
Key Details
- Engineers use Gemini models for code generation. Some DeepMind employees have also been allowed to use Anthropic's Claude Code, causing internal tension (per BI).
- Pichai reported that a complex code migration — done by agents and engineers working together — was completed six times faster than with engineers alone a year ago.
- Google is shifting to “truly agentic workflows” with more autonomous tasks. Some engineers have specific AI usage goals that factor into 2026 performance reviews.
Context: Industry Trend
Other big tech companies are moving similarly:
- Microsoft CEO Satya Nadella said in April 2025 that 20–30% of code for some projects was AI-written. CTO Kevin Scott predicted 95% of code would be AI-generated within five years.
- Meta set a goal that 55% of code changes in some orgs should be “Agent-Assisted” as of Q4 2025. For H1 2026, 65% of engineers in its creation org are expected to write >75% of committed code using AI.
- Snap said earlier in April 2026 that under its new operating model, at least 65% of new code is AI-generated.
Who This Affects
Developers using AI coding agents (Gemini, Claude Code, Copilot) will see increasing reliance on agentic workflows, with human review remaining mandatory. Performance metrics tied to AI adoption are becoming common at large tech firms.
📖 Read the full source: HN AI Agents
👀 See Also
An Open Standard for Agent Run Records: The Case for a Shared Log Schema
Every agent runtime has its own log format, causing fragmentation in debugging, auditing, and tool portability. The fields already converge on a core schema — it's time to standardize.
The Ghost House Effect: Five Failure Modes in AI-Generated Code
Analysis of developer experiences with Claude AI reveals a consistent pattern: initial 10x productivity gains followed by compounding technical debt. The source identifies five specific failure modes including shadow dependencies and context window paralysis.
Anthropic's Business Strategy: API Revenue Drives Consumer Tier Limitations
Anthropic's consumer subscription tiers operate at a loss, subsidized to build AI mindshare, while their API business generates revenue. The $20 Pro tier is intentionally limited to filter users toward higher-value Max subscriptions.
Claude-Code v2.1.91 adds MCP result persistence, shell execution controls, and multi-line deep links
Claude-Code v2.1.91 introduces MCP tool result persistence override via _meta["anthropic/maxResultSizeChars"] annotation supporting up to 500K characters, adds disableSkillShellExecution setting, and enables multi-line prompts in claude-cli://open?q= deep links with encoded newlines.