Measuring AI Agent Autonomy in Practice

Overview

Anthropic researchers analyzed millions of human-agent interactions across Claude Code and their public API to understand how AI agents are deployed in real-world settings. The study examines autonomy levels, user experience effects, risk profiles, and oversight mechanisms.

Key Findings

Increasing Autonomous Runtime The longest Claude Code sessions have nearly doubled in duration over three months, from under 25 minutes to over 45 minutes at the 99.9th percentile. This steady increase across model releases suggests users are granting more independence over time, not just capability improvements.

Shifting Oversight Patterns As users gain experience, "roughly 20% of sessions use full auto-approve, which increases to over 40% as users gain experience." Paradoxically, interrupt rates also rise from 5% to 9% of turns—experienced users transition from approving each action to monitoring and intervening when needed.

Agent Self-Monitoring Claude Code requests clarification more frequently than humans interrupt it, particularly on complex tasks. This self-initiated oversight complements external safeguards by surfacing uncertainty proactively.

Risk Distribution Most agent actions remain low-risk and reversible. Software engineering represents nearly 50% of agentic activity on the public API, with emerging but limited use in healthcare, finance, and cybersecurity. Only 0.8% of actions appear irreversible.

Research Methodology

The team drew from two complementary data sources:

• Public API: Broad visibility across thousands of customers, analyzing individual tool calls
• Claude Code: Complete session tracking enabling workflow analysis

Both use privacy-preserving infrastructure to maintain user confidentiality while enabling empirical observation.

Recommendations

Model developers should invest in post-deployment monitoring infrastructure and train systems to recognize their own uncertainty. Product developers should design interfaces supporting effective user oversight through visibility and intervention mechanisms, rather than mandating specific approval patterns.

The research emphasizes that effective agent oversight emerges from interaction between model design, user behavior, and product architecture—factors invisible to pre-deployment testing alone.