Anthropic measures AI agent autonomy in real deployments

On February 18, 2026, Anthropic published a study examining how much autonomy AI agents are actually being given in deployed environments. The research uses a privacy-preserving tool to analyze millions of human-agent interactions across Claude Code and Anthropic’s public API. The company’s questions were practical rather than theoretical: how much latitude do people grant agents, how does that change with experience, what domains are agents being used in, and how risky are the actions those agents are taking? The paper’s broader message is that post-deployment monitoring is becoming a necessary part of safe agent rollout.

The headline result is that longer autonomous runs in Claude Code are increasing quickly. Anthropic says the 99.9th percentile turn duration rose from under 25 minutes in October 2025 to over 45 minutes in January 2026. At the same time, the share of full auto-approve sessions climbed with experience: new users use it in roughly 20% of sessions, while users with around 750 sessions exceed 40%. Importantly, that does not mean oversight disappears. Human interrupt rates also rise with experience, suggesting that users are shifting from approving every action to monitoring the agent and stepping in only when redirection is needed.

Anthropic also argues that agent-initiated stopping is an important oversight mechanism. On the most complex tasks, Claude Code asked for clarification more than twice as often as humans interrupted it. On the public API side, software engineering accounted for nearly 50% of agentic activity, but the company also observed emerging usage in healthcare, finance, and cybersecurity. Anthropic says most public API actions it observed were still low-risk and reversible, and that risky deployments exist but are not yet happening at broad scale.

The bigger implication is that there is still a gap between what models may be capable of and how much autonomy they are trusted with in practice. Anthropic characterizes this as a kind of deployment overhang: models may be able to handle more independence than real-world users currently allow. At the same time, the combination of higher auto-approve rates, higher interrupt rates, and more agent-initiated pauses suggests that oversight in the agent era is moving away from step-by-step approval toward selective intervention and better model self-calibration. Anthropic’s conclusion is that safer agent deployment will require stronger post-deployment monitoring infrastructure and new human-AI interaction patterns that treat oversight as a shared system property rather than a single permission toggle.