AI Coding Agents Now Complete 35 Percent of Pull Requests at Major Tech Companies, Internal Data Shows

Internal engineering metrics shared informally across several large technology companies point to a consistent pattern: AI coding agents — not autocomplete, but agents that plan, write, test, and iterate on a task with minimal human intervention — are now responsible for roughly 30 to 35 percent of merged pull requests in some engineering organizations, up from single-digit percentages just twelve months earlier. The shift tracks closely with the release cadence of agentic coding tools through 2026: Claude Sonnet 5's coding-focused agentic capabilities, GitHub Copilot Workspace's expanded autonomous task completion, and Cursor's agent mode all reached general availability or major capability upgrades within the same six-month window. The common thread across the tools is a move from single-file autocomplete to multi-file, multi-step task execution with automated test-writing and self-correction loops. The engineering management response has been to reorganize code review practices rather than review volume. Several engineering leads describe a shift toward reviewing AI-authored pull requests for architectural fit and edge-case handling rather than line-by-line correctness, on the reasoning that syntax and logic errors are now caught by the agent's own test-and-iterate loop before a human ever sees the diff. The metric with the most disagreement among engineering leaders is defect rate. Some report AI-agent-authored code has a lower post-merge defect rate than human-authored code for well-scoped tasks, attributing this to more consistent test coverage generated alongside the code. Others report the defect rate is comparable but the failure mode is different — AI agents rarely make syntax errors but occasionally introduce subtle logic errors in edge cases that pass all written tests but fail on inputs the test suite did not anticipate, a category of bug that requires a different review discipline than what most teams have historically practiced.