Evolving beliefs and reasoning strategies were observed in 22 fifth- and sixth-grade children who worked over 8 weeks for a total of about 5 h on a causal reasoning problem. Children planned, performed, and interpreted experiments to learn about the relations between design features and speed of race cars in a computerized microworld. The group made progress, but by the end of the sessions did not fully understand those features that disconfirmed their initial beliefs. In their activity with the microworld, children often failed to make informative comparisons or valid judgments about the outcomes. Exploratory strategies improved as children exercised them over time, but invalid heuristics that preserved children's favored theories about cars were evident throughout. Those children using more valid strategies achieved more complete, stable comprehension of the microworld's structure. In turn, children used their beliefs to make meaning of the complex patterns of evidence they observed. The most successful children evaluated both the evidence and their changing theories, and were sensitive to the fact that they should be mutually constraining.