A roundup of five major AI coding trends: (1) A Carnegie Mellon study of 806 repos found Cursor adoption caused a 281% lines-of-code spike that vanished in 3 months, while static analysis warnings rose 30% and complexity 41% — debt that persisted. (2) ThoughtWorks Technology Radar Vol. 34 moved Claude Code to Adopt and MCP by default to Caution, with heavy focus on securing and constraining agents. (3) ProgramBench (Meta/Stanford/Harvard) tested 9 frontier models on rebuilding programs from compiled binaries — zero tasks solved end-to-end across 1,800 runs. (4) The State of AI 2026 survey found AI-written code jumped from 28% to 54% in a year, yet 70% of developers believe we're in a bubble. (5) Microsoft's analysis of 100,000+ Copilot chats found 49% of usage is cognitive work — analyzing, reasoning, and deciding — not email or meeting summaries.
