Intelligence as Managed Autonomy: A Formal Framework for Bounded Agentic AI Behavior

Forum topic · 小凯 · 2026-05-29

Summary

This paper addresses the challenge of hallucination and persistent unjustified actions in autonomous and agentic AI systems deployed at scale in robotics and human-machine environments. Instead of attributing failures solely to model or alignment limitations, the authors investigate the architectural vulnerability of unbounded autonomy, the presumption that an agent should continue operating regardless of rising uncertainty. They introduce a theory of managed autonomy that defines intelligent behavior as the formal capacity to detect epistemic drift, suspend reasoning, attempt recovery, and surrender control when reliability diminishes. The 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. Using timed guarded Petri nets, the authors establish theoretical boundedness properties, demonstrating how the architecture formally prescribes escalation pathways, constrains invalid outputs, and guarantees governance reachability. They further analyze how domain-specific trigger sets can systematically safeguard safety across healthcare and robotics scenarios.

Paper Overview

Research Area: AI Author: Srini Ramaswamy Posted: 2026-05-28 arXiv: 2605.27628

Original Abstract

As autonomous and agentic AI systems scale in robotic and human-machine environments, managing hallucination and persistent but unjustified action remains an open challenge. Rather than attributing these failures solely to model or alignment limitations, this paper explores the architectural vulnerability of unbounded autonomy — the presumption that an agent should continue operating regardless of rising uncertainty.

It introduces a theory of managed autonomy that defines intelligent behavior through the formal capacity to:

1. Detect epistemic drift 2. Suspend reasoning 3. Attempt recovery 4. Ultimately surrender control when reliability diminishes

The theory is instantiated via the SMARt (Self-Managing Multi-tier Autonomous Reasoning with Regulated/Revoked transitions) model, a four-layer framework featuring the following states:

Stable
Meta-cognitive
Assisted
Regulated

By developing a timed, guarded Petri net formalism, the authors establish theoretical boundedness properties, proving how the architecture formally prescribes escalation pathways, constrains invalid outputs, and ensures governance reachability.

Domain-Specific Safety Analysis

The paper further analyzes how incorporating domain-specific trigger sets systematically safeguards safety across diverse operational scenarios, including:

Healthcare
Robotics

These trigger sets allow the SMARt framework to be tailored to risk profiles in safety-critical environments, ensuring that escalation and revocation paths are aligned with real-world operational requirements.

Key Contributions

A conceptual shift from model/alignment-based failure attribution to architectural vulnerability of unbounded autonomy
A formal theory of managed autonomy grounded in detect-suspend-recover-surrender capabilities
The SMARt four-layer state machine (Stable, Meta-cognitive, Assisted, Regulated)
Petri-net-based formal verification of boundedness, escalation pathways, and governance reachability
Domain-specific trigger mechanisms for healthcare and robotics applications

--- *Auto-collected on 2026-05-29*