← Bibliography

Measuring AI agent autonomy in practice

View original ↗

Note

Primary data: median Claude Code turn  45 seconds; 99.9th percentile grew from ¡25 min to ¿45 min (Oct 2025–Jan 2026). Auto-approve rate grows from 20% for new users to 40%+ for experienced users (750+ sessions). Claude stops to ask clarification 2x more often than users interrupt on complex tasks.

Citation Key

anthropic2026agentautonomy