Show HN: Agentic interface for mainframes and COBOL

A new Show HN post is pitching something that sounds like a sci-fi reboot of office software: an AI helper for the giant old-school computers that still run chunks of banking, government, and other critical systems. In plain English, it promises to help people poke around those famously stubborn text-based screens, inspect files, write job scripts, debug failures, and generally make a decades-old world feel a little less trapped in 1979.

But the real action was in the comments, where the vibe swung wildly between "finally!" and "absolutely not." One camp saw a rescue mission for aging systems nobody wants to touch. A blunt hot take declared that US banks and creditors need this yesterday, which is about as close to a standing ovation as Hacker News gets. Another commenter was cautiously optimistic: even if AI cannot safely crank out flawless finance code on day one, maybe it could at least write decent tests for old COBOL programs, which many readers seemed to treat as surprisingly useful.

Then came the paranoia, jokes, and corporate fear. One commenter immediately asked the question every big company lawyer probably screamed internally: what was the training data? If these systems contain valuable private business logic, nobody wants their secret sauce feeding the machine. The website's promise not to train on customer data helped, but did not end the anxiety. And for the funniest line of the thread, one skeptic compared unleashing an AI on a mainframe to letting a fox into a henhouse. Somewhere between nostalgia, desperation, and dread, the crowd basically agreed on one thing: if this works, it matters. If it fails, it fails in the most dramatic place possible.