New research from Anthropic reveals how people are actually using AI agents in the wild, analyzing millions of interactions across Claude Code and their public API. The study found that Claude Code sessions have nearly doubled in length, with the longest-running sessions growing from under 25 minutes to over 45 minutes in just three months. This trend suggests agents are capable of more autonomy than they typically exercise.

Among experienced users, auto-approval rates jumped from 20% to over 40% as people gained confidence in the system. Interestingly, Claude Code pauses for clarification more often than humans interrupt it, with clarification requests occurring more than twice as frequently on complex tasks. The research also found that while agents are being used in risky domains like healthcare and finance, most API activity remains in low-risk software engineering tasks.

The findings highlight a critical gap between agent capabilities and real-world deployment. As AI agents become more sophisticated, understanding how humans actually interact with and oversee these systems becomes essential for safe deployment. The research points to a need for new monitoring infrastructure and human-AI interaction paradigms that balance autonomy with appropriate oversight.