Anthropic reveals how we actually supervise AI agents

Everyone talks about AI agents taking over human workloads, but how much leash do we actually give them when real code and data are on the line?

Anthropic just released an empirical study analyzing millions of interactions across Claude Code and its public API. The core takeaway: today’s models can handle far more autonomy than users are currently comfortable handing over.

Power users let go, but keep a hand on the brake

For the vast majority of tasks, agent sessions remain short. The median turn duration in Claude Code hovers around 45 seconds, showing little change over the past several months. Most people still use agents as immediate task runners.

Look at the heavy-duty tail end, however, and the picture shifts dramatically. Between October 2025 and January 2026, the 99.9th percentile of continuous autonomous operation nearly doubled, climbing from under 25 minutes to over 45 minutes.

User behavior changes predictably with familiarity. Newcomers with fewer than 50 sessions opt for full auto-approval in only about 20% of their runs. Once users cross 750 sessions, that figure jumps past 40%. Experienced operators stop micro-reviewing every action and let Claude run solo, choosing to intervene only when something looks off.

Interestingly, the agent itself plays hall monitor. On the most demanding tasks, Claude Code pauses to ask for clarification more than twice as often as human users interrupt it. Inside Anthropic, Claude Code's success rate on internal complex tasks doubled between August and December, while average human interventions dropped from 5.4 to 3.3 per session.

Code dominates, high-risk domains stay cautious

Zooming out to the broader API traffic across thousands of enterprise customers, software engineering still drives nearly 50% of all agent tool usage. While agent deployments are beginning to trickle into healthcare, finance, and cybersecurity, virtually all real-world actions remain low-risk and easily reversible.

This gap points to what Anthropic calls a deployment overhang. While external capability evaluations like METR indicate that Claude Opus 4.5 can tackle tasks equivalent to nearly 5 hours of human work in isolated tests, real-world autonomy peaks at around 42 to 45 minutes before human guidance or agent self-checkpoints kick in.

Why it matters

This study provides an essential reality check against pure bench testing. For developers and teams, it shows that expanding autonomy is a matter of building habit and trust rather than waiting for next-generation frontier weights. Watching power users shift to auto-approval at twice the baseline rate confirms that getting comfortable with agent handoffs takes hundreds of real sessions.

For the industry, we believe this clearly shifts the competitive battleground toward oversight and control interfaces. The bottleneck to broader agent adoption is no longer raw intelligence, but the friction of monitoring.

Anthropic's emphasis on agent-initiated stops is a particularly smart design philosophy. The safest path to longer-running autonomy isn't letting an agent drift unattended for hours; it's teaching the agent when to stop and raise its hand. Platforms that master this collaborative safety loop will be the ones that finally unlock agents outside software engineering.