Artificial intelligence is advancing in clear stages, with each stage adding a new capability that is already shaping how physicians work. In clinical practice, these capabilities show up in documentation, image interpretation, decision support, and care coordination.
Training, however, has not been organized around these stages, even though doing so would make adoption more practical and more effective. A better approach is to teach physicians to engage with each layer of AI in the same stepwise way they learn clinical medicine.
The progression of artificial intelligence can be understood through four simple terms that reflect how these capabilities build on one another. LLMs (large language models) focus on generating and understanding language. RAG (retrieval-augmented generation) adds access to external knowledge such as guidelines and datasets. An AI agent uses that information to carry out tasks within a workflow. At the highest level, agentic AI coordinates multiple agents over time to support more continuous and complex processes. Each stage builds on the one before it, creating a layered progression that expands what these tools can do in real-world clinical settings.
READ: Sreedhar Potarazu | Is AI an avatar of God? Anthropic, Mythos, and the rise of moral authority in machines (April 13, 2026)
Ophthalmology provides a useful example because it combines imaging, pattern recognition, longitudinal care, and procedural decision-making. The progression from LLM to RAG, then to AI agent, and finally to agentic AI can be mapped directly onto how an ophthalmologist would learn to use these tools in daily practice.
The first layer involves LLMs. An ophthalmologist encountering this layer would begin with tasks that are already familiar, such as writing clinic notes, summarizing visits, or drafting patient instructions. This is where AI scribes fit naturally into practice. These tools convert conversations into structured documentation using language models. At this stage, the goal is to understand how the output is generated and to maintain full control over clinical reasoning.
For example, after examining a patient with mild non-proliferative diabetic retinopathy, the physician may use an AI scribe to draft an assessment and plan. The result may read clearly and completely, but the physician must confirm that staging, follow-up, and management reflect the actual clinical findings.
The second layer introduces RAG. Here, the ophthalmologist begins to use tools that bring in clinical guidelines, prior cases, and research evidence to support decision-making. In practice, this could involve reviewing a retinal image while also checking current treatment recommendations or comparing findings with similar cases. Tools like OpenEvidence fit directly into this layer by allowing physicians to query medical literature and receive summarized, source-linked answers in real time.
This shifts the physician’s role from recalling evidence to continuously validating decisions against dynamically retrieved knowledge. The key skill at this stage is evaluation. The physician must determine whether the information applies to the patient in front of them, taking into account disease severity, comorbidities, and prior response to treatment.
The third layer moves into the use of an AI agent. At this point, AI begins to participate directly in clinical workflow. An ophthalmologist might use an AI agent that flags glaucoma progression based on visual field and OCT data, suggests additional testing, generates documentation, and initiates follow-up.
READ: Sreedhar Potarazu | Vance vs Rubio round 2 : The Anteambulo strategy (April 18, 2026)
A widely recognized example is Viz.ai, which identifies large vessel occlusions on imaging and alerts care teams in real time (not yet in ophthalmology). Similar approaches are emerging in ophthalmology, where AI agents assist in identifying disease patterns and triggering next steps. At this stage, the physician’s role expands to supervising actions. For example, if an AI agent recommends escalation of glaucoma therapy, the physician must determine whether the change reflects true progression or expected variability.
The final layer involves agentic AI. In this stage, multiple AI agents work together over time to support ongoing care. In ophthalmology, this could include continuous monitoring of patients with glaucoma or diabetic retinopathy, combining imaging analysis, follow-up scheduling, documentation, and patient communication into a coordinated process. An ophthalmologist working at this level uses these capabilities to stay ahead of disease progression while maintaining a full view of the patient’s experience. For instance, a patient may meet clinical criteria for treatment escalation, yet factors such as medication tolerance, adherence, and lifestyle remain central to the decision. Training at this stage focuses on maintaining longitudinal judgment while using these tools to support continuity.
When these layers are taught in sequence, adoption becomes more natural because each step builds on the previous one. The physician begins with LLMs through tools like AI scribes, adds RAG through evidence-based retrieval platforms, learns to supervise AI agents that participate in workflow, and ultimately integrates agentic AI into long-term care. This mirrors how clinical expertise develops, moving from foundational skills to more advanced decision-making over time.
In medical school, this framework could begin with exposure to LLMs during early clinical documentation and case-based learning, where students learn to compare AI-generated notes with their own reasoning to understand how language-based models differ from clinical interpretation.
READ: Sreedhar Potarazu | Caste for Indians in America is still alive (April 10, 2026)
In residency, training would progress to structured use of RAG tools such as OpenEvidence during case discussions and patient workups, with explicit emphasis on validating retrieved evidence against individual patient context. As residents advance, supervised use of AI agents would be introduced in real clinical workflows, such as flagging imaging abnormalities or supporting follow-up planning, with attendings reinforcing when to accept, modify, or override recommendations.
By the final stages of training, exposure to agentic AI would focus on longitudinal care management, where trainees learn how multiple coordinated tools support chronic disease monitoring while they remain responsible for synthesis, prioritization, and patient-centered decision-making. This progression ensures that by the time physicians enter independent practice, they are not simply users of AI tools, but trained supervisors of layered intelligence operating within clinical care.
This approach reinforces what defines clinical judgment. An ophthalmologist does not simply identify findings on a retinal image but connects those findings to disease progression, patient risk, and real-world impact. Vision loss affects independence, daily function, and quality of life, and these considerations shape every decision. Learning to use AI in layers allows physicians to preserve this integrated perspective while taking advantage of new capabilities.
Training aligned with the progression of LLMs, RAG, AI agents, and agentic AI allows physicians to adopt these tools with clarity and confidence. Each layer becomes a practical skill, and each interaction strengthens clinical thinking while expanding what is possible in patient care.