The Irreversibility Budget: Fleet-Level Risk Accounting and Admission Control for Agent Operating Systems

What the source reports

The engineering layer supplies the mechanisms. The Irreversibility Budget (2609.00275) shows why local gates fail at fleet scale: per-effect checks let a fleet of individually-authorized agents overdraw its principal's risk under a shared trigger - in the controlled study, up to 48× the tenant's risk limit. The fix is to treat irreversibility as a first-class metered resource: a trusted runtime charges each effect its residual value-at-risk and denies the marginal effect once the aggregate would overdraw. Spawn Freely, Act Sparingly (2609.01035) extends the same logic to recursive agent trees: distinguish sandbox spawning (external controls prevent the harm) from capability activation (the branch crosses an irreversible-action boundary), and hold a trajectory-level risk budget in escrow, debiting it as branches activate. The paper proves an anytime harm bound and derives an "authority reproduction number" R_A where trajectory harm changes character as it crosses one. Design rule: search broadly in the sandbox; grant recursive authority sparingly, with an explicit risk charge.