I believe there was a change recently to add message history to the context passed to sub-agents. I think I understand why there's upside to that, but it makes for some weird behaviour.
I'll post an example -- essentially, the sub-agent gets confused by the back-and-forth of the messages.
That is, even though it has a system prompt telling it to act as an actual agent and do what the input was to the agent, the last message, it instead duplicated the network routing agent's responses, because that's what it saw in the chain.