Studies of agreement attraction in language production have shown that speakers systematically produce verb agreement errors in the presence of a local noun whose features differ from that of the agreement controller. However, in attraction experiments, these errors only ever occur in a subset of trials. In the present study, we applied a naturalistic scene-description paradigm to investigate how attraction affects the distribution of errors and the time-course of correctly inflected verbs. We conducted our experiment both in the lab and in an unsupervised web-based setting. The results were strikingly similar across the experimental settings for both the error and timing analyses, demonstrating that it is possible to conduct production experiments via the internet with a high level of similarity to those done in the lab. The experiments replicated the basic number attraction effect, though they elicited comparable interference from both singular and plural local nouns, challenging common assumptions about a strong plural markedness effect in attraction. We observed slowdowns before correct verbs that paralleled the distribution of agreement errors, suggesting that the process resulting in attraction can be active even when no error is produced. Our results are easily captured by a model of agreement attraction in which errors arise at the point of computing agreement, rather than reflecting earlier errors made during initial encoding of the subject number.