Researchers have a name for the sinking feeling you get when an AI agent does something you never asked for: delegation regret.
Twenty university students used OpenClaw, a general-purpose AI agent, to complete five everyday tasks chosen to vary in privacy, stakes, and reversibility. The researchers tracked trust, perceived control, transparency, and supervision burden for each task, then coded students' written reflections. Trust didn't move with how high-stakes a task was. It moved with whether the action could be undone and whether someone else would see it: an email task triggered the steepest trust drop (3.10 out of 5) and the strongest demand for approval before sending (4.65 out of 5), while a riskier but easily verified task barely registered.
The more interesting finding is that regret showed up even when the agent's output was rated a success. The problem wasn't bad results, it was acting before showing its work. That reframes agent design as a transparency problem, not just an accuracy one, and it pushes back on the industry's current enthusiasm for agents that run multi-step tasks unattended in the background.
Twenty students and five tasks won't settle anything on their own, but the discomfort will sound familiar to anyone who has watched an agent fire off an email before they had finished reading the draft.