Anthropic publishes research measuring real-world autonomy of AI agents including Claude Code session lengths

Anthropic published research analyzing millions of human-agent interactions across its Claude Code product and public API. The study used a privacy-preserving tool to assess how much autonomy users grant agents in practice. The research found that Claude Code is working autonomously for longer periods. Among the longest-running sessions, the length of time before stopping nearly doubled in three months from under 25 minutes to over 45 minutes. The increase occurred smoothly across model releases. Experienced users auto-approved more frequently. Full auto-approve sessions rose from roughly 20 percent among new users to over 40 percent as users gained experience. On the most complex tasks, Claude Code stopped to ask for clarification more than twice as often as humans interrupted it. Most agent actions on the public API remained low-risk and reversible. Software engineering accounted for nearly 50 percent of agentic activity. Emerging usage appeared in healthcare, finance and cybersecurity, though not yet at scale. Anthropic concluded that effective oversight of agents will require new forms of post-deployment monitoring infrastructure and new human-AI interaction paradigms.