Agent execution trajectories can expand into branching tree structures as subagents collect data from multiple sources before returning a final response