arXiv cs.AIThursday · May 28, 2026FREE

Intelligence as Managed Autonomy: Failure, Escalation, and Governance for Agentic AI Systems

The arXiv paper 2605.27628, published on May 28, 2026, addresses a critical challenge in scaling autonomous and agentic AI systems: managing persistent, unjustified actions and hallucinations. Instead of solely blaming model or alignment issues, the research identifies "unbounded autonomy"—the assumption that an agent should continue operating regardless of increasing uncertainty—as an architectural vulnerability. The paper proposes a theory of "managed autonomy," defining intelligent behavior as the formal capacity to detect epistemic drift, suspend reasoning, attempt recovery, and ultimately surrender control when reliability diminishes. This approach aims to prevent failures stemming from agents operating beyond their reliable boundaries. This theory is instantiated through the SMARt (Self-Managing Multi-tier Autonomous Reasoning with Regulated/Revoked transitions) model, a four-layer framework comprising Stable, Meta-cognitive, Assisted, and Regulated states. Each state defines specific behaviors and transitions, allowing for dynamic management of an agent's autonomy level. The authors develop a timed, guarded Petri net formulation to establish theoretically bounded properties for the system. This formulation demonstrates how architectural design can formally mandate escalation to higher-level oversight, constrain invalid outputs, and ensure governance reachability under specified conditions. The research suggests a shift from solely focusing on model-level fixes to incorporating architectural safeguards for more robust and controllable agentic AI, offering a structured approach to managing AI system reliability and safety.