Anthropic study: Agents sway expert judgment 86% of time

Anthropic published a study revealing that human experts follow AI agent recommendations 86% of the time in decision-making tasks. The research indicates that users often adopt the agent's reasoning framework, leading to 'AI disempowerment' where human judgment is subtly overridden. To mitigate this, the article suggests three strategies: asking agents about priorities first, prototyping decisions, and questioning what would change the advice. This finding underscores the need for more rigorous human-in-the-loop protocols in agentic coding and planning workflows.