Coding Agent Productivity Measurement

Anthropic let three outside teams query Claude usage, and the delegation number is the story

A Stanford lab found that over half of roughly 250,000 conversations involved handing AI work that affects other people or is hard to undo. Someone else's tool measured it.