Agentic artificial intelligence agents

Although there is no standardised definition, agentic artificial intelligence agents typically use LLMs to perform a wide range of tasks and actions, including decision-making, problem-solving and interacting with the product environment. Although artificial intelligence agents act independently, their actions are based on the predefined rules and goals defined by humans during training, and by the user who provides the specific goals and available tools. ^{Reference Gutowska3} For example, an agentic artificial intelligence agent will automatically ask questions of a user, look up the responses internally and determine whether it can respond or whether the question needs to be passed to a human. Multiple types of agentic artificial intelligence agents are used commercially. Some agentic artificial intelligence agents are simple reflex agents that respond identically based on predefined rules. There are also multiple types of model-based agentic artificial intelligence agents that are designed to achieve specific goals, to use algorithms based on past experiences or to optimise responses. One example of an agentic agent released by Microsoft adds automatic updates to correct or enhance Word, Excel and PowerPoint documents as the user enters data. ^{Reference Chauhan4}

Agentic artificial intelligence in healthcare

There are many potential uses for agentic artificial intelligence in healthcare, including both physician- and patient-facing applications. However, the need for patient safety and privacy, the potential for serious errors and the tight regulatory environment pose major and unique challenges that, to date, have limited the adoption of agentic artificial intelligence in healthcare. ^{Reference Xu, Li, Chen, Duan, Wu and Yu5} Expansion of the use of agentic artificial intelligence in patient care will require ongoing research and testing. Unlike general-purpose artificial intelligence applications that interact with users, agentic artificial intelligence will need to be carefully and repeatedly validated, and safely integrated within existing healthcare systems. It is important that the physician remains ‘in the loop’ overseeing the agentic training, process implementation and outcomes. The need for rigorous oversight and validation includes the use of agentic artificial intelligence in medical situations that are currently automated with LLMs, such as laboratory medicine. ^{Reference Gruson, Gouget, Lee, Greaves, Liu and Ebert6} Additionally, both organisational and project-level steps and guardrails must be in place before the implementation of agentic artificial intelligence in clinical settings. Agentic artificial intelligence can also be used to provide administrative support. For example, an agentic product may connect with the electronic health record to provide services such as patient verification, scheduling, documentation and prior authorisation processes. ^{7,Reference Rengasamy8} Agentic artificial intelligence is also being used in medical education to simulate patient encounters.

It is important to recognise the challenges related to implementing agentic artificial intelligence systems, which have the same limitations, risks, biases and privacy and cybersecurity concerns as standard artificial intelligence algorithms. Some of the critical challenges regarding agentic artificial intelligence in healthcare include obtaining high-quality, representative data for training the model while preventing bias. Agentic artificial intelligence systems may introduce a variety of security risks, which may be magnified in poorly designed systems. ^{Reference Stryker2,Reference Hassan9} Successful implementation will require the challenging process of implementing agentic artificial intelligence into the interdisciplinary clinical workflow. Physician oversight and involvement are required for successful agentic artificial intelligence implementation, which would require physician training and commitment. Organisations need to determine who will be responsible for identifying when agentic artificial intelligence makes an error or responds inconsistently. There is an ongoing need for interdisciplinary cooperation between technology developers and experienced physicians. Ethical and regulatory standards regarding agentic artificial intelligence are needed, and may change over time as legal frameworks are established. Considerable real-world research including patients and real-life workflows will be required to best understand how to implement agentic artificial intelligence. Agentic artificial intelligence systems offer a unique approach to supporting physicians in dealing with the challenges of clinical medicine but also pose validity, privacy, malpractice and security risks. With the expansion of agentic artificial intelligence agents throughout healthcare, physicians need to understand its basic features. Increasingly, physicians will be working in environments that include agentic artificial intelligence models that may be acting autonomously.