“Agent-ception”

Like everyone else, I have been playing along with AI Agents, though I do my best to not let the agents do my thinking for me and instead treat them like an intern.
Recently I observed a phenomenon worth noting:
Here is an exact quote from a coding session:
The chain makes sense now — my sub-coder spawned one nested sub-coder, which apparently spawned another itself (invisible to the middle agent).
My agent spawned another agent, which spawned yet another agent.
Then the agents doing the spawning of other agents were not even aware of their subagents.
Besides this sounding like a great way to burn dollars on tokens, it also raises some serious infosec concerns.
If an agent can’t even keep track of its sub-agents and their sub-agents how are we, the human in the loop, supposed to keep track of this whole thing?
It’s also funny in a dark way. An agent thinks, “Wow, this is a lot of work. I’ll just pass the buck to the next guy”, then the second agent thinks the same and so on.
That might make a good comic, but the phenomenon isn’t well known enough to resonate with people.
Hopefully the harness will be able to track both agent lineage and limit how many nested sub-agents can be spawned.
It doesn’t seem like a monumental task.
In the meantime, you will likely see a dystopian game of telephone being played by your agents and their sub-agents that is, at best, inefficient and a waste of tokens and at worst, a possible infosec disaster waiting to happen.
Have you observed anything like this?