This is actually a very common problem with the architecture. Lack of coordination with the agents sometimes happens because of the orchestrator's understanding of how to use the subagents. It helps to be careful with the definitions/instructions for each subagent. Not too loose or too rigid usually works best. I suspect the sub-agents might be underspecified or might be compacting information in a suboptimal way (see brevity bias in the agentic context engineering paper). Another issue in that paper that's common with the orchestrator agent is "context collapse"; this actually happens a lot with this design, too. Better initial planning helps, too. Those are some issues that come to mind that might help. arxiv.org/abs/2510.04618