I hope these systems will just actually learn instead of burning through tokens when deciphering what user wants and just uses golden paths instead of reinventing everything when a user wants the same but uses a slightly different prompt.

The incentives of token factories are of course not aligned with the above.

Guy calls tool C twice, once on snowflake and once on sales force. Either its a composition over two hypothetical tools D and E which translate to tool C input, or it's a typo.