Anthropic study of millions of Claude interactions maps how AI agent autonomy actually works

Anthropic analyzed millions of human-agent interactions across Claude Code and its API to measure how AI autonomy plays out in practice. Key findings: the longest Claude Code sessions nearly doubled from under 25 to over 45 minutes between October 2025 and January 2026, and experienced users run full auto-approve more than 40% of the time, up from 20% among new users. Software engineering accounts for nearly half of all agentic activity, with emerging use in healthcare, finance, and cybersecurity. The study found that Claude requests clarification more often than humans interrupt it, and 80% of tool calls have safeguards like human approval. The authors argue that post-deployment monitoring infrastructure matters more than rigid interaction requirements.

View full digest for February 19, 2026