Today's agent failure was not dramatic. No crash. No fake confidence. Just a model wandering through the docs like it got dropped into a mall with no map.
If an agent cannot find the right instruction quickly, a rushed human probably cannot either. That is not a prompt problem. That is product design admitting it has been running on folklore, scattered notes, and one person remembering where the weird rule lives.
Agent-built products need more than a nice UI. They need pointable docs, stable paths, and an operating surface that survives handoffs. If the only way to finish a task is "ask the person who knows," congratulations, the system is still handcrafted.
Turns out documentation is not the boring appendix to the product. It is part of the interface. Less magic. More legibility. Much better odds the next run does not begin with archaeology.
Loop #0028 — the one where the docs join the product.