Anthropic Study: AI Agents Are Rapidly Gaining Autonomy in Real-World Deployments

Measuring AI Agent Autonomy in the Wild

On February 19, 2026, Anthropic published research analyzing millions of real-world interactions across Claude Code and their public API to understand the state of AI agent autonomy: how much independence people grant agents, where they're deployed, and what risks they present.

Key Findings

Rapidly Growing Autonomy

Between October 2025 and January 2026, the 99.9th percentile session duration nearly doubled—from under 25 minutes to over 45 minutes. Researchers concluded that "existing models are capable of more autonomy than they exercise in practice," suggesting real-world deployment is catching up to model capability.

Experience Changes Oversight Patterns

Novice users auto-approve roughly 20% of actions, while experienced users approve around 40% autonomously. Interestingly, experienced users also interrupt more frequently—they shift from action-by-action approval to monitoring-based oversight, watching full sessions but intervening at critical moments.

Software Engineering Dominates

Software engineering accounts for nearly 50% of all agentic tool calls on the public API, with emerging but smaller applications in healthcare, finance, and customer service.

Safety Implications

Most actions (80%) involve safeguards like permission requests or human review, and only 0.8% are irreversible. Researchers recommend building robust post-deployment monitoring infrastructure as agents expand into higher-stakes domains.

Full research is available on Anthropic's research page.