Tiered Agentic Oversight: A Hierarchical Multi-Agent System for Healthcare Safety
Yubin Kim, Hyewon Jeong, Chanwoo Park, Eugene Park, Haipeng Zhang, Xin Liu, Hyeonhoon Lee, Daniel McDuff, Marzyeh Ghassemi, Cynthia Breazeal, Samir Tulebaev, Hae Won Park
TL;DR
This paper presents Tiered Agentic Oversight (TAO), a hierarchical multi-agent system designed to improve safety in healthcare AI applications by layered supervision, error correction, and collaboration with human doctors, outperforming existing systems on safety benchmarks.
Contribution
The paper introduces TAO, a novel hierarchical multi-agent framework inspired by clinical hierarchies, enhancing safety and error correction in healthcare AI systems.
Findings
TAO absorbs up to 24% of individual agent errors.
TAO outperforms single-agent systems on 4 out of 5 safety benchmarks with up to 8.2% improvement.
Lower tiers are essential; their removal significantly reduces safety.
Abstract
Large language models (LLMs) deployed as agents introduce significant safety risks in clinical settings due to their potential for error and single points of failure. We introduce Tiered Agentic Oversight (TAO), a hierarchical multi-agent system that enhances AI safety through layered, automated supervision. Inspired by clinical hierarchies (e.g., nurse-physician-specialist) in hospital, TAO routes tasks to specialized agents based on complexity, creating a robust safety framework through automated inter- and intra-tier communication and role-playing. Crucially, this hierarchical structure functions as an effective error-correction mechanism, absorbing up to 24% of individual agent errors before they can compound. Our experiments reveal TAO outperforms single-agent and other multi-agent systems on 4 out of 5 healthcare safety benchmarks, with up to an 8.2% improvement. Ablation studies…
Peer Reviews
Decision·Submitted to ICLR 2026
1. The TAO framework, inspired by clinical hierarchies, presents a clear and innovative conceptual contribution to AI safety. 2. It demonstrates strong empirical performance, outperforming both single-agent and multi-agent baselines on most healthcare safety benchmarks.
1. The paper lacks large-scale experiments, such as scaling to more agents or additional tiers. 2. The framework incurs higher computational costs, as indicated in Table 2. 3. The central routing mechanism may become a bottleneck, and the analysis of its robustness is limited. 4. Defining clear scopes for each tier is challenging, which could lead to overlapping responsibilities or gaps between agents.
1. The hierarchical structure with adaptive escalation is a thoughtful adaptation of clinical workflows to multi-agent AI. It provides a clear and structured mechanism for error correction and oversight. 2. The paper includes experiments across multiple benchmarks that assess safety, accuracy, and ethical alignment. 3. The clinician-in-the-loop study adds practical credibility, showing that the framework can work effectively with human oversight in real-world settings.
1. While the hierarchical agent design is interesting, it largely builds on existing multi-agent paradigms such as debate or voting frameworks. The paper does not clearly articulate how its approach advances beyond prior work or contributes new theoretical insights. 2. The experiments are mainly healthcare-focused and lack sufficient details for reproducibility. Key materials such as code are not provided, and important aspects like statistical significance are missing. The user study is small i
- Authors focus on a timely and novel topic by proposing a medical safety-focused multi-agent framework that tries to address a critical gap in healthcare safety. - The main results of the paper in Table 2 demonstrate strong performance of TAO, where it outperforms other ingle-agent, multi-agent, and adaptive systems across five different benchmarks. - Authors also conduct comprehensive ablation studies including but not restricted to different adversarial setups, attribution analysis, tier conf
- The main limitation of the proposed framework the practical usability in the wild due to its high computational cost. As shown in Table 2, TAO results with the highest cost, sometimes nearly double the previous state-of-the-art. Since paper presents a more application-oriented solution with MAS rather than a theoretical or algorithmic improvements, cost is a critical concern. Thus, this raises two questions: (1) is the slight accuracy gain of up to ~4% (sometimes 0.04%, 0.7%, etc.) meaningful
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsHealthcare Operations and Scheduling Optimization · Occupational Health and Safety Research