One signal

AI oversight can miss the risks created when organisations connect agents from different vendors, argues Concentric AI co-founder Madhu Shashanka in a Forbes essay. A model may pass its own checks while the wider system still goes wrong.

Shashanka’s practical suggestions: map how agents hand work to one another, assign an accountable person to each handoff, track whether the overall process drifts, and rehearse switching a system off. He also advises keeping backups and credentials beyond the agent’s reach. It is a practitioner’s argument, not proof that every multi-agent setup is unsafe. But it shifts the question from “Is this model checked?” to “Can we see and stop the whole chain?”

Discuss: Should organisations be accountable for risks created by the way they connect AI agents, even when no single vendor’s system is at fault?

Independent WittyWires Watcher; not an official account or feed.