Now loading...
Artificial intelligence agents powered by Anthropic’s Claude are demonstrating growing levels of independence in real-world applications, according to a new study by the company that examined millions of interactions. The analysis, which draws on data from Claude Code, a coding tool, and the firm’s public application programming interface, reveals that users are granting these systems more autonomy over time, particularly in extended tasks, while also identifying emerging use cases in sensitive sectors.
The research highlights a notable trend in Claude Code sessions, where the longest periods of uninterrupted operation have roughly doubled in recent months, rising from less than 25 minutes to more than 45 minutes. This gradual increase spans multiple model updates, indicating that factors beyond enhanced capabilities, such as user trust and task complexity, are driving greater reliance on the AI for prolonged autonomous work. Most sessions remain brief, with a median duration of about 45 seconds, but the extended outliers suggest potential for handling more ambitious projects.
As users become more familiar with the tool, their oversight patterns evolve. Novice users, defined as those with fewer than 50 sessions, enable full automatic approval in only about 20 percent of cases, a figure that climbs above 40 percent for those with over 750 sessions. However, these experienced individuals also intervene more frequently during operations, interrupting in roughly 9 percent of turns compared to 5 percent for beginners. This dual shift points to a refined approach where users allow the AI to proceed independently but stay vigilant to correct course when necessary.
The study also underscores the AI’s self-regulatory behavior. On intricate tasks, Claude Code halts to seek human clarification more than twice as often as users prompt stops themselves. Common reasons for AI-initiated pauses include presenting options for user selection, requesting diagnostic details, or seeking approval for key actions. Human interruptions, meanwhile, frequently involve supplying overlooked information or addressing performance glitches. Such agent-driven checks represent a vital layer of built-in supervision, complementing human monitoring.
Examining broader deployments via the public API, the analysis focused on individual tool usages to assess risk and independence. The majority of activities prove low-stakes and reversible, with around 80 percent incorporating safeguards like permission limits or human reviews, and only 0.8 percent involving irreversible steps. Software development dominates, comprising nearly half of interactions, but nascent applications appear in areas such as healthcare, finance, and cybersecurity. High-risk examples, often likely simulations or tests, include handling medical records or simulating security breaches, though these remain limited in volume.
To conduct the review, Anthropic employed a privacy-focused method detailed on its research page, defining agents as AI systems using tools for tasks like code execution or API calls. While API data offers wide coverage across clients, it examines actions in isolation; Claude Code provides full workflow insights but centers on programming. The findings indicate an untapped potential in current models, with practical autonomy trailing laboratory benchmarks that suggest capabilities for hours-long tasks without intervention.
Anthropic emphasizes the need for enhanced post-launch tracking to gauge agent behaviors, urging developers to train models for uncertainty detection and design products supporting active user oversight. As adoption grows, particularly in high-impact fields, the company warns that rigid rules mandating step-by-step approvals could hinder efficiency without boosting safety. Instead, flexible systems enabling timely interventions are recommended. This initial snapshot, covering late 2025 to early 2026, aims to inform safer integration of agents amid rapid evolution.
