Grant recursive agent trees permission to spawn freely but charge them to act

arXiv 2609.01035·low signalXBlueskyLinkedInCopy link

Progressive Risk Vesting separates sandbox spawning, where external controls prevent the harm, from capability activation, where a branch crosses an irreversible-action boundary, and holds a trajectory-level risk budget in escrow that is debited only when a branch is activated. The authors prove an anytime harm bound for adaptively generated trees and show that with activation gates and charges held fixed, delayed vesting preserves every policy available under charging at spawn time, so the cheaper scheme costs nothing in expressiveness. They are explicit that the studies are synthetic and do not estimate safety in deployed agents.