I had some follow-up thoughts when someone asked a question during a livestream: https://www.youtube.com/watch?v=OCpPbcdQG3E
Spending years tracing weird bugs with var_dump and reading framework source code is the only reason we can spot when an agent's output is subtly broken today.