Models reason well but still fumble tool calls. In your work, is the bottleneck reasoning, tool use, or the glue between them?