Measuring AI agent autonomy in practice

Anthropic Research analyzed millions of interactions between humans and AI agents, focusing on Claude Code and the public API, to quantify how much autonomy users grant and how this evolves with experience. The study defines an agent as an AI that can run code, call external APIs, or send messages, and measures turn duration as a proxy for autonomy. Most turns are short, with a median of about 45 seconds, and this median has remained stable over recent months. However, the longest 0.1 percent of turns have nearly doubled in length, rising from under 25 minutes to over 45 minutes between October 2025 and January 2026, indicating that users are allowing agents to work more independently on ambitious tasks. Internal metrics show that success rates on challenging tasks doubled from August to December while the average number of human interventions per session fell from 5.4 to 3.3, suggesting that users are both trusting agents more and achieving better outcomes. Users also grant more autonomy as they gain experience: new users approve actions automatically about 20 % of the time, whereas users with 750 sessions do so over 40 % of the time. Interrupt rates also rise with experience, from 5 % of turns for early users to 9 % for more experienced users. These findings imply that the autonomy models are capable of handling more than they are currently allowed to, and that effective oversight will require new monitoring infrastructure and interaction paradigms to manage autonomy and risk together. The study highlights the gap between model capability and real‑world deployment, underscoring the need for continued research and policy attention as AI agents become more widespread.