(1) Given any task or workflow,
(2) With any amount of steps,
(3) Written by any person in any format and in any language,
(4) Across any API integration(s),
(5) With any mix and amount of structured and unstructured data...
Our models need to:
(1) Understand what the user wants the outcome to be
(2) Know what level of quality is actually satisfactory
(3) One-shot the entire workflow that creates that outcome
While also:
(1) Costing less than the value the outcome creates
(2) Not taking forever
(3) Not hallucinating
(4) Being able to do it again. And again. And again.

^^ this is the Agent problem we've been struggling with.