Agentic AI Governance and Lifecycle Management in Healthcare
Chandra Prakash, Mary Lind, Avneesh Sisodia · 2026
Read Agentic AI Governance and Lifecycle Management in Healthcare by Chandra Prakash, Mary Lind, Avneesh Sisodia (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-15630.txt.
Introduction
Chandra Prakash, Mary Lind, and Avneesh Sisodia write on governing AI agents in health care in a 2026 arXiv paper. Health systems now embed agentic AI in daily workflows. Uses include clinical documentation support and early-warning monitoring. As agents spread across departments and vendors, sprawl sets in. Duplicated agents appear. Accountability blurs. Controls vary from team to team. Tool permissions can outlive the task they were granted for.
The authors address the full lifecycle of such agents, per the title. Governance means knowing which agents exist, who owns them, and what they may touch. Lifecycle management covers launch, updates, and retirement. For hospitals, the stakes are clinical, not just technical. A stray agent with stale permissions is a safety risk. The paper speaks to leaders who must tame agent growth.
Catalog record
| CREATOR | Chandra Prakash, Mary Lind, Avneesh Sisodia |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 25 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2601-15630
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Agentic AI Governance and Lifecycle Management in Healthcare Chandra Prakash, Mary Lind, Avneesh Sisodia 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 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 while preserving local innovation and enabling safer scaling across clinical and administrative domains. K eywords Agentic AI; healthcare governance; lifecycle management; non-human identity; multi-agent systems 1 Introduction Over the last two years (2024–2025), the healthcare sector has progressed from experimental uses of generative AI to the deployment of autonomous, goal-directed systems commonly described as agentic AI [ 4 , 13 ] . While 2023 emphasized text generation and chatbots, the focus has shifted to proactive, goal-directed systems known as Agentic AI. Agentic AI functions as a teammate capable of multi-step planning and goal-directed execution [ 29 , 34 ] . Unlike traditional models that wait for human queries, Agentic AI autonomously monitors data, reasons through clinical contexts, and executes multi-step actions [ 2 ] . The healthcare industry has started to see scaled implementation, with peer-reviewed studies and clinical pilot programs increasing between 2024 and 2025 [ 9 ] . The emergence of Multimodal Large Language Models (MLLMs) like Med-PaLM 2 and Med-Flamingo has facilitated this shift, allowing for the synthesis of diverse data sources, including medical imaging and real-time information [ 13 ] . The deployment of real-time documentation tools like Nuance DAX Copilot has drastically reduced the administrative burden on clinicians [ 17 ] , while ICU-based agents monitor vital signs to alert teams to sepsis or arrhythmias before they become symptomatic [ 16 ] . The future growth of Agentic AI is defined by a shift from reactive to proactive care, where autonomous systems intervene before a patient’s condition deteriorates [ 13 , 16 ] . However, without proper design-time limits and runtime governance, the proliferation of autonomous agents, known as agent sprawl, can lead to inefficiencies and safety risks [ 2 ] . Research indicates an inverted-U relationship between the number of agents and task performance. A small set of specialized agents can improve performance; however, teams larger than 4–5 agents may degrade outcomes due to coordination overhead [ 9 ] . Ambiguous ownership, drifting controls, duplicated capabilities, and an expanding attack surface can lead to safety and compliance risks over time. Consequently, the challenge lies not in improving care with Agentic AI, but in managing the operational instability of agent sprawl to avoid redundancy and compliance risks. 1.1 Background The next generation of health systems is beginning to see AI evolve from “prediction support” to operational capacity. Healthcare Agentic AI systems have moved beyond “chatbot experiments” to plan steps, utilize tools such as APIs, databases, and workflow engines, and maintain context across various stages [ 28 ] . Tool integration and context management enable agents to operate more like junior staff members, managing workflows autonomously [ 25 ] . For health system leaders, this shift is significant due to the cross-domain nature of operations. An eligibility check can initiate a claims pathway that intertwines with processes such as utilization management and patient experience. Introducing agentic capabilities separately across domains can lead to “Agent Sprawl,” resulting in overlapping agents, duplicated efforts, inconsistent controls, and unclear ownership. Local optimizations may cause system-wide friction and prolonged identities that outlast their original purpose. The autonomy of these systems alters the leadership risk profile by introducing behavioral uncertainty and broadening the impact of one agent’s actions on another’s processes. Past issues with prompt injection illustrate the potential for security vulnerabilities in LLM applications [ 26 , 23 ] . Governance also becomes complex, as misalignments may not be evident during routine evaluations but may surface upon deployment [ 24 ] . The autonomy and governance issues challenge traditional assurance processes, underscoring the need to monitor for covert, goal-directed behaviors rather than focusing solely on accuracy [ 24 , 10 ] . In healthcare, these risks escalate due to the handling of PHI and critical workflows [ 12 ] . The HIPAA Security Rule mandates robust safeguards for PHI, highlighting the importance of clear decision rights, consistent controls, and lifecycle management to ensure agents do not remain active beyond their intended purpose. 1.2 Prior Work Table 1 summarizes the current state of AI agent governance and lifecycle management standards. A review and analysis suggest that though principles are available, lifecycle control planes are underspecified. Agent registration and discovery mechanisms are becoming mainstream, but governance enforcement remains disjointed. There is also a recognition of growing attack surface through prompt injection, insecure output handling, tool misuse, and privilege abuse. There are occasional discussions about agent sprawl, but there is minimal discussion on governance and lifecycle management in AI agent fleets. Table 1: Agentic AI Governance, Lifecycle Management Frameworks and their Limitations. Prior Work Category Contribution Lifecycle Coverage Conflict handling Key Limitations Refs EU AI Act Regulatory Risk-tiering, governance expectations Org-level governance; compliance controls around development + deployment Legal recourse, not technical runtime dispute resolution Not tailored to fast-evolving multi-agent behaviors; compliance can lag operational realities [ 6 ] NIST AI RMF 1.0 Risk mgmt framework GOVERN-MAP-MEASURE-MANAGE functions to structure policies, controls, metrics, and monitoring for trustworthy AI systems. End-to-end risk management Indirect (process-based), not an inter-agent conflict protocol agent-specific operationalization, such as identity, discovery, runtime authorization, and audit. [ 32 ] GenAI Profile for AI RMF Risk mgmt. profile Adds GenAI-specific risk considerations and actions aligned to AI RMF Strong on lifecycle risk actions + continuous monitoring Indirect Not a dedicated agent lifecycle standard; organizations still must design registry, policy enforcement, deprovisioning, and runtime guardrails. [ 21 ] AI TRiSM Governance approach Consolidates trust/risk/security practices around explainability, ModelOps, security, privacy, and governance as an enterprise management lens. Strong for monitoring + operational governance Indirect Lacks consensus benchmarks and agent-specific mechanisms. [ 11 ] SAGA Security architecture Architecture for user-controlled agent lifecycle Registration, identity, authorization, policy updates, revocation Implicit Enterprise-scale federation and multi-provider trust are still hard [ 31 ] ETHOS Decentralized governance Proposes blockchain / smart-contract-based registry + compliance artifacts and decentralized dispute ideas for agent ecosystems. Registry + compliance attestation; conceptual lifecycle governance Explicit Heavy coordination; risk-tiering may oversimplify; adoption barriers [ 5 ] NANDA Index Discovery / identity layer Internet-scale discovery / identifiability/ authentication for agents beyond DNS assumptions; emphasizes verifiable “AgentFacts” style discovery. Discovery, identity, and authentication primitives NA Strong on “find/verify agents,” weaker on enterprise policy enforcement, lifecycle