Type: url
Original: https://www.anthropic.com/research/measuring-agent-autonomy
Status: generated
Summary
Anthropic’s February 2026 study analyzes millions of human-agent interactions across Claude Code and its public API to measure how much autonomy people grant AI agents in practice. It finds that the longest Claude Code turns nearly doubled in length between October 2025 and January 2026, that experienced users both auto-approve and interrupt more often, that Claude Code stops to ask for clarification more often than humans interrupt it on complex tasks, and that most public-API agent actions are low-risk, with software engineering accounting for nearly half of tool calls. It closes with recommendations on post-deployment monitoring, training models to recognize their own uncertainty, and designing products for user oversight.