Similar to log, this can be used to pass along extra information about what exact
messages were predicted by the LLM before parsing out the (tool, tool_input).
This is again useful if (tool, tool_input) cannot be used to fully recreate the
LLM prediction, and you need that LLM prediction (for future agent iteration).
Compared to log, this is useful when the underlying LLM is a chat model (and
therefore returns messages rather than a string).
