Measuring AI agent autonomy in practice
Summary
Anthropic analyzes how AI agents operate in the real world by measuring autonomy across Claude Code and its public API. The study finds that autonomy grows over time, with users granting more freedom while also interrupting more often, and that Claude Code frequently pauses for clarification. The authors argue for post-deployment monitoring and new human-AI interaction paradigms to manage autonomy and risk, noting frontier domains where actions are riskier and less supervised.