# How People Actually Use AI Agents Source: https://www.youtube.com/watch?v=WBN1CyBy8bE Recap page: https://rapidrecap.app/video/WBN1CyBy8bE Generated: 2026-02-19T22:33:11.564+00:00 --- ## Quick Overview The Anthropic study "Measuring AI agent autonomy in practice" reveals that real-world AI agents exhibit significantly less autonomy than previously assumed, with users frequently interrupting or asking for clarification, especially on more complex tasks, indicating that capability improvements alone do not fully explain observed autonomy. **Key Points:** - Real-world AI agents currently have much less autonomy than often assumed, as users frequently interrupt or ask for clarification, particularly on complex tasks. - The 99.9th percentile turn duration for Claude Code nearly doubled from under 25 minutes (late September 2025) to over 45 minutes (early January 2026), indicating increasing autonomy usage despite model releases. - Users interrupt Claude code for clarification (asking for options) more often than they interrupt for intervention (stopping work), especially as task complexity rises. - For high-complexity tasks, humans interrupt Claude 8.8% of the time, while Claude asks for clarification 16.4% of the time. - The primary reason Claude stops itself is to present the user with a choice (35%), contrasting with the top human interruption reason: providing missing context/corrections (32%). - Software engineering accounts for nearly 50% of all agentic tool calls observed, showing its current dominance in deployment. - The study suggests that increased autonomy is due to a combination of factors, including user trust building and applying Claude to increasingly ambitious tasks, not purely model capability increases. ![Screenshot at 00:00: The title screen for the Anthropic blog post "Measuring AI agent autonomy in practice" sets the context for the discussion about real-world agent usage and control.](https://ss.rapidrecap.app/screens/WBN1CyBy8bE/00-00-00.jpg) **Context:** This video summarizes findings from Anthropic's February 18, 2026, study titled "Measuring AI agent autonomy in practice," which investigates how users actually interact with and supervise AI agents like Claude Code. The study contrasts theoretical capability assessments with real-world data derived from user interactions, focusing on turn durations, interruption rates, and clarification requests to quantify the actual autonomy exhibited by these agents. ## Detailed Analysis The Anthropic study on AI agent autonomy in practice finds that real-world agents possess significantly less autonomy than anticipated, heavily relying on human interaction. The research analyzed millions of interactions across the public API and Claude Code. A key finding is the increasing turn duration for the most ambitious users (99.9th percentile), which nearly doubled from under 25 minutes in late September 2025 to over 45 minutes in early January 2026, suggesting growing complexity in agent tasks rather than just model capability improvements. Furthermore, agent-initiated stops (clarifications) occur more frequently than human interruptions, especially as task complexity increases. For high-complexity tasks, humans interrupt 8.8% of the time, whereas Claude requests clarification 16.4% of the time. The most common reason for agent-initiated stops is presenting the user with choices (35%), while the top human interruption reason is providing missing context or corrections (32%). The study also highlights that software engineering accounts for almost 50% of all agentic tool calls, indicating its current primary deployment domain. The overall trend suggests that increased autonomy stems from users building trust and applying models to harder tasks, rather than solely from raw model performance gains. Finally, a key takeaway is that autonomy is not just about steps taken (capability) but also about permission scope and the ability to change state (human interaction context). ### Introduction & Autonomy Definition - AI agent autonomy is not just steps taken, but permission scope and ability to change state - Agent is an AI system equipped with tools to take actions like running code, calling external APIs, and sending messages. ### Turn Duration Trends - 99.9th percentile turn duration nearly doubled from 25 minutes to over 45 minutes between late September 2025 and early January 2026 - This increase is smooth across model releases, suggesting factors beyond model capability. ### Human Intervention vs. Clarification - Users interrupt for clarification more often than for intervention, especially on complex tasks - For high complexity, humans interrupt 8.8% of turns, while Claude asks for clarification 16.4% of turns. ### Reasons for Stopping - Top agent-initiated stop is presenting user with choice (35%) - Top human interruption reason is providing missing context/corrections (32%) - Humans interrupt less frequently (around 5%) when they are experienced. ### Deployment Domains - Software engineering accounts for nearly 50% of all tool calls - Back-office automation is second at 9.1% - Other domains are below 10%. ![Screenshot at 00:00: The title screen for the Anthropic blog post "Measuring AI agent autonomy in practice" sets the context for the discussion about real-world agent usage and control.](https://ss.rapidrecap.app/screens/WBN1CyBy8bE/00-00-00.jpg) ![Screenshot at 00:37: A chart from METR showing AI ability to complete long tasks, illustrating the rapid improvement in capability over time.](https://ss.rapidrecap.app/screens/WBN1CyBy8bE/00-00-37.jpg) ![Screenshot at 00:50: A graph illustrating the 99.9th percentile turn duration for Claude Code, showing a steady increase from under 25 minutes to over 45 minutes.](https://ss.rapidrecap.app/screens/WBN1CyBy8bE/00-00-50.jpg) ![Screenshot at 01:41: A transition slide showing the core concept of measuring autonomy by tracking time elapsed between start and stop/interruption.](https://ss.rapidrecap.app/screens/WBN1CyBy8bE/00-01-41.jpg) ![Screenshot at 02:30: A table summarizing common reasons why Claude stops itself versus why humans interrupt it, showing differing priorities.](https://ss.rapidrecap.app/screens/WBN1CyBy8bE/00-02-30.jpg)