Abstract
<title>Abstract</title> <p>Artificial Intelligence (AI) agents are increasingly being explored for automation purposes in healthcare applications. However, their deployment raises concerns regarding the reliability, usability, and trustworthiness in clinical decision support. In particular, Large Language Model (LLM) based systems often exhibit variability between different runs on the same input, which makes them unsuitable for safety-critical clinical applications. To address these challenges, we proposed an Agentic OODA Framework, a multi-agent-based implementation of the Observe–Orient–Decide–Act (OODA) decision-making paradigm for struc-tured clinical decision support. The framework breaks clinical reasoning into modular, stateful agents that take part in observation, contextual orientation , automatic decision-making and validation with human-in-the-loop enabled cycles. By integrating similarity-based self-consistency checks and persistent execution states, our multi-agent framework supported a structured and traceable decision making process. We evaluated the proposed framework in the context of clinical decision support for stroke and report generation and measured the reliability of the framework in terms of the consistency and stability of the output generated when the same input was executed 5 times for 30 patients. The Agentic OODA Framework was found to be significantly more reliable with less variation in the explanatory text output with a mean cosine similarity of 0.95 when compared with other general LLM models such as Claude Sonnet 4.5, Claude Haiku 4.5, GPT 5.1, GPT 5.2, Llama 4 Maverick, Gemini 3 Pro Preview, Gemini 3 Flash, Qwen 3-32B and DeepSeek V3.2. Feedback loops such as human verification further guarantees patient trust and clinical safety. These results indicated that the use of LLM-based decision-making, which is structured and controlled within the agentic OODA framework, is likely to significantly improve reliability, 1 auditability, and clinical appropriateness. The proposed framework offers a gen-eralizable solution for applying AI agents in healthcare settings, which require high levels of consistency and transparency.</p>