Do experienced AI users trust agents more or just oversee differently?

Anthropic's research report measures how Claude Code users change their oversight behavior as they gain experience with the tool, and finds two numbers rising together. On auto-approval: "newer users (<50 sessions) employ full auto-approve roughly 20% of the time; by 750 sessions, this increases to over 40% of sessions." On interruption: "new users (those with around 10 sessions) interrupt Claude in 5% of turns, while more experienced users interrupt in around 9% of turns."

Anthropic names this "an apparent contradiction" and reads it as "a shift in users' oversight strategy" rather than a simple increase or decrease in trust. New users "are more likely to approve each action before it's taken, and therefore rarely need to interrupt Claude mid-execution." Experienced users instead "let Claude work autonomously, stepping in when something goes wrong or needs redirection" — oversight moves from pre-approval to monitoring-and-correction. Anthropic adds that the higher interrupt rate "may also reflect active monitoring by users who have more honed instincts for when their intervention is needed," and expects the per-turn rate to "eventually plateau," though it says widening confidence intervals at high session counts make that hard to confirm yet. A parallel pattern appears on Anthropic's public API, where "87% of tool calls on minimal-complexity tasks...have some form of human involvement, compared to only 67% of tool calls for high-complexity tasks" — which Anthropic itself calls "counterintuitive."

This complicates Can we design AI autonomy separately from capability?, whose taxonomy treats a user's role — operator, approver, observer, and so on — as a design choice a developer fixes for a given agent. Anthropic's data shows the same user moving across that spectrum as their own experience changes, without any change to the agent's design: toward something closer to approver or observer for routine actions, while retaining an operator-like interrupt capacity for the moments that need it. It also qualifies Does granting agents more autonomy undermine human oversight?: that paper frames extended use of agent systems as something that erodes the judgment oversight depends on, but Anthropic's experienced users interrupt more often, not less — consistent with sharpened rather than only degraded instincts, at least on this measure, though a rising interrupt rate against a shrinking share of reviewed actions does not rule out that each intervention now catches a smaller slice of total behavior.

The finding is Anthropic's own interpretation of its own product's telemetry, not a controlled study: the claim that experience changes oversight strategy, rather than that more experienced users simply select different or more trustworthy tasks, is offered as a probable reading ("may also reflect") rather than a demonstrated mechanism, and whether the interrupt rate truly plateaus is explicitly unconfirmed in the source.