pith. sign in

arxiv: 2601.15630 · v2 · pith:LIGC4AJDnew · submitted 2026-01-22 · 💻 cs.AI

Agentic AI Governance and Lifecycle Management in Healthcare

Chandra Prakash , Mary Lind , Avneesh Sisodia This is my paper

Pith reviewed 2026-05-21 15:14 UTC · model grok-4.3

classification 💻 cs.AI
keywords Agentic AIHealthcare governanceLifecycle managementAI oversightControl planesAgent securityCompliance frameworks
0
0 comments X

The pith

A five-layer blueprint called UALM gives healthcare leaders concrete controls for managing fleets of agentic AI systems.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Healthcare groups are adding agentic AI to daily tasks such as documentation and monitoring, which quickly creates duplicated agents, unclear ownership, and permissions that outlive their purpose. The paper synthesizes governance standards, security practices, and compliance rules into the Unified Agent Lifecycle Management blueprint. This blueprint divides oversight into five control planes that handle identity registration, cross-domain coordination, data boundaries, runtime rules with kill switches, and full decommissioning tied to audit trails. If the approach works, leaders can keep local experimentation while gaining consistent, auditable oversight that supports broader rollout.

Core claim

The authors derive UALM from a rapid synthesis of existing governance standards, agent security literature, and healthcare compliance needs. The blueprint organizes recurring operational gaps into five control-plane layers: an identity and persona registry, orchestration and cross-domain mediation, PHI-bounded context and memory, runtime policy enforcement with kill-switch triggers, and lifecycle management linked to credential revocation and audit logging. A companion maturity model guides staged adoption. On the authors' terms, UALM supplies an implementable pattern that delivers audit-ready oversight while allowing continued local innovation and safer scaling in both clinical and non-clin

What carries the argument

The Unified Agent Lifecycle Management (UALM) blueprint, which maps agent governance gaps onto five layered control planes that together cover registration, coordination, data scoping, enforcement, and decommissioning.

If this is right

  • CIOs and CISOs obtain a ready pattern for registering every agent and revoking its credentials at end of life.
  • Runtime enforcement with kill switches can stop agents from acting outside approved PHI contexts.
  • Orchestration layers reduce duplication by mediating access across departments and vendors.
  • The maturity model lets organizations begin with basic identity controls and add layers over time.
  • Audit logging tied to decommissioning produces traceable records for compliance reviews.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same layered structure might transfer to other regulated sectors that face agent sprawl, such as finance or logistics.
  • Hospitals could test the blueprint by retrofitting it onto a small set of existing agents and tracking changes in oversight effort.
  • Without vendor-specific implementation guides, adoption may vary widely depending on the technical maturity of each health system's IT stack.

Load-bearing premise

That combining existing standards and literature in a rapid synthesis will reveal and cover all important recurring gaps without direct testing in operating healthcare environments.

What would settle it

A controlled pilot in one health system that measures the count of untracked agents, permission persistence, and audit completion rates before and after applying the five UALM layers for six months.

Figures

Figures reproduced from arXiv: 2601.15630 by Avneesh Sisodia, Chandra Prakash, Mary Lind.

Figure 1
Figure 1. Figure 1: Visualization of Agentic AI Maturity Model [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Agentic AI Governance and Lifecycle Management Framework [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The Unified Agent Lifecycle Management (UALM) Reference Architecture. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
read the original abstract

Healthcare organizations are beginning to embed agentic AI into routine workflows, including clinical documentation support and early-warning monitoring. As these capabilities diffuse across departments and vendors, health systems face agent sprawl, causing duplicated agents, unclear accountability, inconsistent controls, and tool permissions that persist beyond the original use case. Existing AI governance frameworks emphasize lifecycle risk management but provide limited guidance for the day-to-day operations of agent fleets. We propose a Unified Agent Lifecycle Management (UALM) blueprint derived from a rapid, practice-oriented synthesis of governance standards, agent security literature, and healthcare compliance requirements. UALM maps recurring gaps onto five control-plane layers: (1) an identity and persona registry, (2) orchestration and cross-domain mediation, (3) PHI-bounded context and memory, (4) runtime policy enforcement with kill-switch triggers, and (5) lifecycle management and decommissioning linked to credential revocation and audit logging. A companion maturity model supports staged adoption. UALM offers healthcare CIOs, CISOs, and clinical leaders an implementable pattern for audit-ready oversight that preserves local innovation and enables safer scaling across clinical and administrative domains.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper proposes a Unified Agent Lifecycle Management (UALM) blueprint for healthcare organizations deploying agentic AI. It identifies agent sprawl issues (duplicated agents, unclear accountability, persistent permissions) and derives a five-layer control-plane framework from a rapid synthesis of governance standards, agent security literature, and healthcare compliance requirements. The layers are: (1) identity and persona registry, (2) orchestration and cross-domain mediation, (3) PHI-bounded context and memory, (4) runtime policy enforcement with kill-switch triggers, and (5) lifecycle management and decommissioning linked to credential revocation and audit logging. A companion maturity model for staged adoption is also presented, with the claim that UALM provides an implementable pattern for audit-ready oversight while preserving local innovation.

Significance. If the proposed mapping and layers can be shown to close the identified operational gaps, the work would offer practical value to healthcare CIOs, CISOs, and clinical leaders by supplying a structured, audit-oriented governance pattern for agent fleets. The synthesis of existing standards with agent-specific controls is a constructive contribution, but the absence of any validation, traceability, or pilot evidence substantially limits the current significance.

major comments (3)
  1. [Abstract] Abstract: The central claim that the five control-plane layers comprehensively address recurring gaps (duplicated agents, persistent permissions, unclear accountability) and deliver audit-ready oversight rests on an unvalidated rapid synthesis. No detailed gap inventory, traceability matrix to source standards/literature, or empirical evidence (pilot, retrospective analysis, or simulation) is supplied to show that the layers actually mitigate these problems in operational healthcare settings.
  2. [UALM Blueprint] The UALM blueprint description: The assertion that the layers (identity registry, orchestration, PHI-bounded context, runtime enforcement, lifecycle decommissioning) map onto the gaps in a manner sufficient for safer scaling is presented as a direct output of the synthesis, yet the manuscript contains neither the explicit mapping details nor any test against real agent-sprawl cases that would substantiate the comprehensiveness claim.
  3. [Maturity Model] Maturity model section: The companion maturity model is introduced to support staged adoption, but no specific criteria, measurable indicators, or examples are provided to demonstrate how adoption stages would produce verifiable improvements in oversight or risk reduction.
minor comments (2)
  1. [Abstract] The phrase 'rapid, practice-oriented synthesis' would benefit from a brief description of the literature scope, time frame, and selection criteria to allow readers to assess completeness.
  2. [UALM Blueprint] Notation for the five layers could be made more consistent (e.g., numbered list with short labels) to improve readability when the layers are referenced later in the text.

Simulated Author's Rebuttal

3 responses · 1 unresolved

We thank the referee for their thoughtful review and constructive criticism. We have carefully considered each major comment and provide point-by-point responses below. Where appropriate, we will revise the manuscript to incorporate additional details and clarifications.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central claim that the five control-plane layers comprehensively address recurring gaps (duplicated agents, persistent permissions, unclear accountability) and deliver audit-ready oversight rests on an unvalidated rapid synthesis. No detailed gap inventory, traceability matrix to source standards/literature, or empirical evidence (pilot, retrospective analysis, or simulation) is supplied to show that the layers actually mitigate these problems in operational healthcare settings.

    Authors: We acknowledge that the current manuscript does not include a detailed gap inventory or traceability matrix, which would strengthen the presentation of the synthesis. In the revised version, we will add an appendix or dedicated subsection that inventories the identified gaps from agent sprawl and provides a traceability matrix mapping each UALM layer to relevant governance standards, agent security literature, and healthcare compliance requirements. Regarding empirical evidence, as this is a conceptual framework paper proposing a blueprint derived from synthesis, we do not have pilot or simulation data. We will explicitly state this as a limitation and outline plans for future empirical validation. revision: partial

  2. Referee: [UALM Blueprint] The UALM blueprint description: The assertion that the layers (identity registry, orchestration, PHI-bounded context, runtime enforcement, lifecycle decommissioning) map onto the gaps in a manner sufficient for safer scaling is presented as a direct output of the synthesis, yet the manuscript contains neither the explicit mapping details nor any test against real agent-sprawl cases that would substantiate the comprehensiveness claim.

    Authors: We agree that explicit mapping details are essential for substantiating the claims. The revised manuscript will include a detailed mapping table that explicitly links each of the five layers to the specific operational gaps (e.g., how the identity and persona registry addresses duplicated agents and unclear accountability). This mapping will be supported by references to the synthesized sources. While we cannot provide tests against real-world cases in this version, the mapping will clarify the rationale behind the layer design. revision: yes

  3. Referee: [Maturity Model] Maturity model section: The companion maturity model is introduced to support staged adoption, but no specific criteria, measurable indicators, or examples are provided to demonstrate how adoption stages would produce verifiable improvements in oversight or risk reduction.

    Authors: We recognize the need for more concrete details in the maturity model. In the revision, we will expand this section to include specific criteria and measurable indicators for each maturity stage, such as the presence of audit logs, percentage of agents with revoked credentials upon decommissioning, and examples of risk reduction metrics. We will also provide illustrative scenarios showing improvements in oversight at different adoption levels. revision: yes

standing simulated objections not resolved
  • The lack of empirical evidence such as pilot studies or simulations to validate the framework's effectiveness in real healthcare settings, as the work is based on a synthesis of existing literature and standards rather than original empirical research.

Circularity Check

0 steps flagged

No circularity: UALM is an external synthesis with no self-referential derivations or fitted inputs

full rationale

The manuscript proposes UALM as a blueprint explicitly derived from a rapid synthesis of pre-existing governance standards, agent security literature, and healthcare compliance requirements. No equations, parameters, or derivations appear in the provided text. The five control-plane layers are presented as a mapping output of that synthesis rather than as quantities fitted to or defined by the paper's own results. No self-citations are invoked as load-bearing uniqueness theorems, and the central claim does not reduce by construction to any internal input. This is a standard conceptual framework paper whose derivation chain is self-contained against external sources.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The proposal rests on the assumption that literature synthesis alone can produce a complete and implementable control framework; no free parameters are fitted, but the five layers and maturity model are constructed for this paper.

axioms (1)
  • domain assumption Existing AI governance frameworks provide limited guidance for day-to-day operations of agent fleets in healthcare.
    This premise is invoked to justify the need for the new UALM blueprint.
invented entities (1)
  • Unified Agent Lifecycle Management (UALM) blueprint no independent evidence
    purpose: To map recurring governance gaps onto five control-plane layers for agentic AI oversight.
    A newly proposed construct without independent evidence or external validation mentioned.

pith-pipeline@v0.9.0 · 5722 in / 1248 out tokens · 78936 ms · 2026-05-21T15:14:11.586861+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

35 extracted references · 35 canonical work pages

  1. [1]

    Brodeur, Thomas A

    Peter G. Brodeur, Thomas A. Buckley, Zahir Kanjee, Ethan Goh, Evelyn Bin Ling, Priyank Jain, Stephanie Cabral, Raja-Elie Abdulnour, Adrian D. Haimovich, Jason A. Freed, Andrew Olson, Daniel J. Morgan, Jason Hom, Robert Gallo, Liam G. McCoy, Haadi Mombini, Christopher Lucas, Misha Fotoohi, Matthew Gwiazdon, Daniele Restifo, Daniel Restrepo, Eric Horvitz, J...

    work page 2025

  2. [2]

    Next-generation agentic ai for transforming healthcare.Informatics and Health, 2:73–83, 9 2025

    Nalan Karunanayake. Next-generation agentic ai for transforming healthcare.Informatics and Health, 2:73–83, 9 2025

    work page 2025

  3. [3]

    Nikahd, Pengfei Dong, Xianqi Li, and Linxia Gu

    Yasin Shokrollahi, Jose Colmenarez, Wenxi Liu, Sahar Yarmohammadtoosky, Matthew M. Nikahd, Pengfei Dong, Xianqi Li, and Linxia Gu. Recent advances in generative ai for healthcare applications. 8 2025

    work page 2025

  4. [4]

    Kamel Boulos

    Peng Zhang and Maged N. Kamel Boulos. Generative ai in medicine and healthcare: Promises, opportunities and challenges.Future Internet, 15:286, 8 2023

    work page 2023

  5. [5]

    Agentic ai in healthcare: A comprehensive survey of foundations, taxonomy, and applications

    Shruti Banerjie, Yuxin Zhu, Isaac Freeman, Julyssa Villa Machado, Abdulaziz Ahmed, Abeed Sarker, and Mohammed Al-Garadi. Agentic ai in healthcare: A comprehensive survey of foundations, taxonomy, and applications. 11 2025

    work page 2025

  6. [6]

    Ai agents in clinical medicine: A systematic review, 8 2025

    Alon Gorenshtein, Mahmud Omar, Benjamin S Glicksberg, Girish N Nadkarni, and Eyal Klang. Ai agents in clinical medicine: A systematic review, 8 2025

    work page 2025

  7. [7]

    Hetherington, Ajay Dharod, Tracey Carroll, Richa Bundy, Hieu Nguyen, Henry E

    Tsai-Ling Liu, Timothy C. Hetherington, Ajay Dharod, Tracey Carroll, Richa Bundy, Hieu Nguyen, Henry E. Bundy, McKenzie Isreal, Andrew McWilliams, and Jeffrey A. Cleveland. Does ai-powered clinical documentation enhance clinician efficiency? a longitudinal study.NEJM AI, 1, 11 2024

    work page 2024

  8. [8]

    A foundational architecture for ai agents in healthcare.Cell Reports Medicine, 6:102374, 10 2025

    Fei Liu, Yue Niu, Qihua Zhang, Kai Wang, Zheyi Dong, Io Nam Wong, Linling Cheng, Ting Li, Lian Duan, Kun Li, Gen Li, Tai Wa Hou, Manson Fok, Huiyan Luo, Xiangmei Chen, Kang Zhang, and Yun Yin. A foundational architecture for ai agents in healthcare.Cell Reports Medicine, 6:102374, 10 2025. 9 APREPRINT- JANUARY23, 2026

    work page 2025

  9. [9]

    Toolformer: Language models can teach themselves to use tools

    Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language models can teach themselves to use tools. 2 2023

    work page 2023

  10. [10]

    Owasp top 10 for large language model applications | owasp foundation

    OW ASP. Owasp top 10 for large language model applications | owasp foundation

    work page

  11. [11]

    Hybrid real-time framework for detecting adaptive prompt injection attacks in large language models.Journal of Computing Theories and Applications, 3:286–301, 1 2026

    Chandra Prakash, Mary Lind, and Elyson De La Cruz. Hybrid real-time framework for detecting adaptive prompt injection attacks in large language models.Journal of Computing Theories and Applications, 3:286–301, 1 2026

    work page 2026

  12. [12]

    An ai-powered coding tool wiped out a software company’s database, then apologized for a ‘catastrophic failure on my part’, 12 2025

    Beatrice Nolan. An ai-powered coding tool wiped out a software company’s database, then apologized for a ‘catastrophic failure on my part’, 12 2025

    work page 2025

  13. [13]

    Detecting and reducing scheming in ai models, 12 2025

    OpenAI. Detecting and reducing scheming in ai models, 12 2025

    work page 2025

  14. [14]

    Bowman, and Evan Hubinger

    Ryan Greenblatt, Carson Denison, Benjamin Wright, Fabien Roger, Monte MacDiarmid, Sam Marks, Johannes Treutlein, Tim Belonax, Jack Chen, David Duvenaud, Akbir Khan, Julian Michael, Sören Mindermann, Ethan Perez, Linda Petrini, Jonathan Uesato, Jared Kaplan, Buck Shlegeris, Samuel R. Bowman, and Evan Hubinger. Alignment faking in large language models. 12 2024

    work page 2024

  15. [15]

    Summary of the hipaa security rule, 12 2009

    HHS. Summary of the hipaa security rule, 12 2009

    work page 2009

  16. [16]

    Ai act, 12 2025

    European Commission. Ai act, 12 2025

    work page 2025

  17. [17]

    Artificial intelligence risk management framework (ai rmf 1.0)

    Elham Tabassi. Artificial intelligence risk management framework (ai rmf 1.0). Technical report, National Institute of Standards and Technology, 1 2023

    work page 2023

  18. [18]

    Artificial intelligence risk management framework :

    NIST. Artificial intelligence risk management framework :. Technical report, National Institute of Standards and Technology, 7 2024

    work page 2024

  19. [19]

    Adib Habbal, Mohamed Khalif Ali, and Mustafa Ali Abuzaraida. Artificial intelligence trust, risk and security management (ai trism): Frameworks, applications, challenges and future research directions.Expert Systems with Applications, 240:122442, 4 2024

    work page 2024

  20. [20]

    Saga: A security architecture for governing ai agentic systems

    Georgios Syros, Anshuman Suri, Jacob Ginesin, Cristina Nita-Rotaru, and Alina Oprea. Saga: A security architecture for governing ai agentic systems. 8 2025

    work page 2025

  21. [21]

    Decentralized governance of autonomous ai agents

    Tomer Jordi Chaffer, Charles von Goins, Bayo Okusanya, Dontrail Cotlage, and Justin Goldston. Decentralized governance of autonomous ai agents. 1 2025

    work page 2025

  22. [22]

    Beyond dns: Unlocking the internet of ai agents via the nanda index and verified agentfacts

    Ramesh Raskar, Pradyumna Chari, John Zinky, Mahesh Lambe, Jared James Grogan, Sichao Wang, Rajesh Ranjan, Rekha Singhal, Shailja Gupta, Robert Lincourt, Raghu Bala, Aditi Joshi, Abhishek Singh, Ayush Chopra, Dimitris Stripelis, Bhuwan B, Sumit Kumar, and Maria Gorskikh. Beyond dns: Unlocking the internet of ai agents via the nanda index and verified agent...

    work page 2025

  23. [23]

    The agntcy agent directory service: Architecture and implementation

    Luca Muscariello, Vijoy Pandey, and Ramiz Polic. The agntcy agent directory service: Architecture and implementation. 9 2025

    work page 2025

  24. [24]

    Announcing the agent2agent protocol (a2a), 12 2025

    Rao Surapaneni, Miku Jha, Michael Vakoc, and Todd Segal. Announcing the agent2agent protocol (a2a), 12 2025

    work page 2025

  25. [25]

    Selective disclosure for agent discovery and identity management (sd- agent)

    Suhas Nandakumar and Cullen Fluffy Jennings. Selective disclosure for agent discovery and identity management (sd- agent). Technical report, Internet Engineering Task Force, 10 2025. Work in Progress

    work page 2025

  26. [26]

    Characterizing ai agents for alignment and governance

    Atoosa Kasirzadeh and Iason Gabriel. Characterizing ai agents for alignment and governance. 4 2025

    work page 2025

  27. [27]

    Ai rmf core, 2025

    NIST. Ai rmf core, 2025

    work page 2025

  28. [28]

    Wico Mulder and André Meyer-Vitali.A Maturity Model for Collaborative Agents in Human-AI Ecosystems, pages 328–335. 2023

    work page 2023

  29. [29]

    Agentic ai in 6g software businesses: A layered maturity model

    Muhammad Zohaib, Muhammad Azeem Akbar, Sami Hyrynsalmi, and Arif Ali Khan. Agentic ai in 6g software businesses: A layered maturity model. 8 2025

    work page 2025

  30. [30]

    Navigating the eu ai act: implications for regulated digital medical products.npj Digital Medicine, 7:237, 9 2024

    Mateo Aboy, Timo Minssen, and Effy Vayena. Navigating the eu ai act: implications for regulated digital medical products.npj Digital Medicine, 7:237, 9 2024

    work page 2024

  31. [31]

    The eu artificial intelligence act (2024): Implications for healthcare

    Hannah Kolfschooten and Janneke Oirschot. The eu artificial intelligence act (2024): Implications for healthcare. Health Policy, 149:105152, 11 2024

    work page 2024

  32. [32]

    Renato Cuocolo, Diana Bernardini, Daniel Pinto dos Santos, Michail E. Klontzas, Tugba Akinci D’Antonoli, Luís Curvo Semedo, Robin Decoster, Merel Huisman, Elmar Kotter, Luis Martí-Bonmatí, Costin Minoiu, Emanuele Neri, Konstantin Nikolaou, Maija Radzina, Evis Sala, Susan C. Shelmerdine, Laurens Topff, and Michelle C. Williams. Ai medical device post-marke...

    work page 2025

  33. [33]

    Ai governance: a systematic literature review.AI and Ethics, 5:3265–3279, 6 2025

    Amna Batool, Didar Zowghi, and Muneera Bano. Ai governance: a systematic literature review.AI and Ethics, 5:3265–3279, 6 2025. 10 APREPRINT- JANUARY23, 2026

    work page 2025

  34. [34]

    William Walter Finch and Marya Butt. Gaps in ai-compliant complementary governance frameworks’ suitability (for low-capacity actors), and structural asymmetries (in the compliance ecosystem)—a systematic review.Journal of Cybersecurity and Privacy, 5:101, 11 2025

    work page 2025

  35. [35]

    Security of llm-based agents regarding attacks, defenses, and applications: A comprehensive survey.Information Fusion, 127:103941, 3 2026

    Yaxin Tang, Yijia Liu, Jiahe Lan, Zheng Yan, and Erol Gelenbe. Security of llm-based agents regarding attacks, defenses, and applications: A comprehensive survey.Information Fusion, 127:103941, 3 2026. 11

    work page 2026