Spawn Freely, Act Sparingly: Progressive Risk Vesting for Recursive LLM-Agent Trees
A paper titled 'Spawn Freely, Act Sparingly: Progressive Risk Vesting for Recursive LLM-Agent Trees' was published on arXiv (cs.AI) on 2026-09-01. It introduces Progressive Risk Vesting (PRV), which holds a trajectory-level risk budget in escrow and debits it as branches are activated. The paper proves an anytime harm bound for adaptively generated trees and analyzes trajectory harm in relation to the authority reproduction number R_A and local risk p.
The paper proposes Progressive Risk Vesting (PRV) for recursive LLM-agent trees, distinguishing sandbox spawning from capability activation. PRV holds a trajectory-level risk budget in escrow and debits it as branches are activated, providing an anytime harm bound for adaptively generated trees. The analysis shows that delayed vesting preserves all policies available under irrevocable spawn charging, and trajectory harm changes as the authority reproduction number R_A crosses one; as local risk p approaches zero, trajectory harm is proportional to p below criticality and proportional to sqrt(p) above it.
PRV introduces a risk-budget escrow mechanism for recursive agent trees, with an anytime harm bound that holds even when branch outcomes are dependent, provided local certificates remain valid conditional on the full pre-activation history. The paper identifies a phase transition in trajectory harm at R_A = 1, with harm scaling linearly in p below criticality and as sqrt(p) above it, suggesting that controlling the authority reproduction number is crucial for safety.
This research addresses a key operational risk in deploying recursive LLM agents: when to grant irreversible action authority. By formalizing risk vesting, it provides a principled approach for enterprises building multi-agent systems to manage tool-use permissions and contain potential harm, potentially enabling safer autonomous workflows.
PRV offers a mechanism to safely scale recursive agent systems by bounding worst-case harm while preserving policy flexibility. This could reduce liability and increase trust in autonomous agent deployments, particularly in high-stakes enterprise applications.
Observable next signals include follow-up work applying PRV to real-world agent frameworks, empirical validation of the harm bound in production systems, and adoption of risk-vesting concepts in agent orchestration platforms or safety standards.