Exposing Limitations of Language Model Agents in Sequential-Task Compositions on the Web