Do people who use AI agents a lot actually trust them more — or do they just learn to watch them differently?

This explores how heavy users of AI agents change the way they supervise them as they gain experience. The short answer from the corpus: they don't simply trust more. They move their checking to a different point in the process. Anthropic's usage data on Claude Code shows that experienced users auto-approve about 40% of sessions, compared with 20% for newer users, yet they also interrupt the agent about twice as often Do experienced AI users trust agents more or just oversee differently?. Newer users act as gatekeepers who approve each step before it runs. Experienced users act more like supervisors who let the agent work and step in when something looks wrong.

A close study of developers shows what this looks like in practice. Experienced users rarely review everything. They rely on shortcuts: passing tests stand in for correctness, a sensible plan stands in for working code, and skimming is the main way they review Do developers actually review agent work exhaustively?. That isn't laziness. Exhaustive review doesn't scale, so people learn where to look. The research-agent work points the same way: having a human step in only at high-uncertainty decision points did better than both full autonomy and step-by-step approval. Constant check-ins turn into rubber-stamping Does targeted human oversight beat both full autonomy and exhaustive review?.

Here's the catch. A 48-person study of computer-use agents found that people blocked only 21% of risky actions once those actions reached execution, and that rate stayed the same whichever oversight approach they used. What actually reduced problems was reviewing the plan up front, which cut problematic actions by 76% Does oversight strategy prevent risky agent actions or just catch them?. So the strategy experienced users drift toward, catching problems mid-run, depends on the step where humans are weakest. The study tested oversight strategies, not experience levels, so it doesn't show that experienced users are worse at catching risky actions. It does suggest the mid-run style feels more reliable than it is. Privacy research adds a similar warning: of six distinct patterns people showed when reviewing agent responses, five showed little awareness that private information was leaking What privacy oversight patterns emerge when people review AI agent responses?. An eye-tracking study shows a related effect in search. Attention moved heavily onto AI Overviews, while trust ratings stayed equal between the overviews and the ranked results Where do searchers look when AI Overviews appear?. Where people look can change without their trust being recalibrated to match.

The design response is to make the experts' instinct work better. Cocoa lets users revise the plan and reassign steps to themselves or the agent while work is underway. That combines up-front plan review with mid-run steering, and it improved control without adding effort Can users steer AI agents better by revising plans mid-execution?. At the organizational level, Deloitte describes the same shift as governance: humans placed at deliberately designed handoff points, backed by tamper-proof logs of what the agent did. That replaces relying on whoever happens to be watching How should organizations supervise autonomous AI agents?.

What you might not expect: in the studies here, the most effective oversight comes before the agent acts, at the plan, while the habit experienced users build is to watch and interrupt. The best setups combine the two by pairing a quick plan check with targeted interruptions at the risky moments.