Humanized agents are often implemented in the user interfaces of computers as well as other devices (e.g., IKEA’s animated online assistant ‘Anna’ and the iPhone speech-based personal assistant ‘Siri’). This study explores the design of humanized agents that interact with respondents in a Web survey questionnaire. Specifically, the agent intervenes when respondents speed, i.e., answer too fast to provide high quality data. Although a few previous studies on this type of intervention reported positive impact on response quality, the evidence also suggests that some respondents may lack the motivation to comply with the intervention requests. We speculate that the intervention could be more effective if respondents treat the intervention as if it is from a human agent rather than simply feedback from the computer; the risk is this may introduce social desirability bias. To explore this, we examine four designs of the intervention for the speeding; in all cases respondents will be prompted and asked to re-consider their answers when their answers are faster than the pre-specified threshold. We manipulate the picture and the text to either be more human-like or feel more like a computer program. We also contrast how the intervention is presented – either on the subsequent survey screen or in a pop-up window – to test the hypothesis that the latter might help create the sense of an agent stepping into the data collection process; as a result, respondents might be more likely to respond socially to the intervention when it involves humanized cues. This study compares the four intervention designs on: 1) respondents’ compliance with the intervention, 2) their subsequent response effort in the survey, and 3) the social desirability of their answers to subsequent sensitive questions.