Problem
Research questions and friction points this paper is trying to address.
governability
enterprise AI
risk-adaptive execution
agent frameworks
computational resource allocation
Beyond Autonomy: A Dynamic Tiered AgentRunner Framework for Governable and Resilient Enterprise AI Execution
🤖 AI Summary
This work addresses the critical gap in existing large language model agent frameworks, which prioritize autonomy at the expense of governability required in enterprise settings—evidenced by unchecked high-risk operations, absent validation for complex tasks, and undifferentiated resource allocation irrespective of risk levels. To bridge this gap, we propose AgentRunner, a novel dynamic hierarchical multi-agent framework that introduces a risk-driven mechanism for adaptive allocation of computational resources and review intensity. The architecture enforces a four-way separation of duties—proposal, review, execution, and verification—and incorporates physical isolation alongside a verification-recovery feedback loop to achieve intrinsic resilience. Empirical evaluation on a real-world multi-tenant SaaS platform demonstrates that AgentRunner simultaneously enhances safety for high-risk tasks and improves resource efficiency, thereby delivering both controllability and robustness essential for enterprise-grade AI deployment.
📝 Abstract
Current large language model agent frameworks prioritize autonomy but lack the governability mechanisms required for enterprise deployment. High-risk write operations proceed without independent review, complex tasks lack acceptance verification, and computational resources are allocated uniformly regardless of risk level. We propose the Dynamic Tiered AgentRunner, a controlled execution protocol distilled from a production-grade multi-tenant SaaS platform. The framework introduces three core mechanisms: (1) Risk-Adaptive Tiering that dynamically allocates computational resources and review intensity based on task risk profiles, achieving Pareto-optimal trade-offs between safety and efficiency; (2) Separation of Powers architecture where proposal, review, execution, and verification are performed by independent agents with physically isolated boundaries; and (3) Resilience-by-Design through a Verifier-Recovery closed loop that treats failure as a first-class system state. We formalize the tier selectio
Innovation
Methods, ideas, or system contributions that make the work stand out.
Risk-Adaptive Tiering
Separation of Powers
Resilience-by-Design
Governable AI
Agent Framework
🔎 Similar Papers
