Notes from the AI Prompt Engineering Masterclass @ Microsoft Burlington, MA, on 26^thMarch, 2025.

I attended my first health-tech focused meetup on the 26^th, and these are my notes from the event.

Prior to the meetup, I did a little bit of pre-reading to understand what to expect, and to have some smart questions to ask if the opportunity came up. Paul, one of the organizers of the community, recently published a fantastic explainer about agentic AI systems, how they work, and how they are relevant to healthcare. An excerpt:

For decades, healthcare has relied on traditional software code designed to follow predefined rules, execute specific tasks, and provide structured outputs based on hardcoded logic. These applications have been essential in digitizing healthcare operations, from EHR systems and scheduling platforms to billing software and ERP systems.

But here's the catch: Traditional software is declarative. It operates exactly as programmed, no more, no less.

Take EMR systems as an example. They're highly structured, allowing clinicians to efficiently input, store, and retrieve patient data. However, if workflow inefficiencies arise—for example, physicians spending too much time on documentation—the software doesn't recognize the problem or suggest a fix. Instead, IT teams must intervene, analyze usage patterns, manually adjust settings, update rules, or integrate third-party solutions to improve performance.

- Paul J. Swider (https://x.com/pswider/status/1896232968693157976)

Agentic AI transforms rule-following systems into thinking systems, alleviating the cognitive overload from an already overburdened, often under-staffed health-system. Going with an example from Paul's post which I quoted above, instead of a system only logging that a hospital is at capacity, agentic AI-enabled systems might reroute admissions, move staff shifts, and make accommodations to adjust to the demand dynamically.

How are such systems orchestrated?

AI Agents typically use LLMs as the decision making layer. Typically, the "execution plan" involves deciding:

What needs to be done
How it needs to be done
Who/what needs to be involved
Tools that must be utilized

Notes from the event and talks

As always, garbage in, garbage out.
Give specific instructions to the model describing the task and the expected output.
Explore Chain-of-thought, role-playing, maieutic prompting: ask the model to answer a question with an explanation, and then follow-up with asking it to explain the explanations. In this way, we can produce tree-of-thought reasoning.

Questions about agentic systems and the influence of prompting frameworks on their effectiveness

If LLMs form the "routing/orchestration layer" of agentic systems, what effect does prompting style have on outcomes? For example, if one node in a graph is a decision node, asking, "does the patient require immediate attention?" is it possible that the way the prompt is written for this node/step causes different outcomes for the same case?
How often do LLMs stray from the desired output (Deep Research agents might not respect the max on search queries or sources to look up, and GPT-4o-mini can often stray from the expected JSON schema when returning answers as structured objects, for example). What kinds of guardrails should developers consider when building agents and do SDKs provide any of them inbuilt?
There are some studies that show that LLMs used in recommendation tasks often display bias on the basis of name, gender, etc. [workplace promotion example]. What kinds of biases do healthcare-specific LLMs like Google's MedPaLM, Microsoft's catalogue of models, etc. show? For reference, we know that patient outcomes are often biased against females due to lack of data from clinical trials, etc. When we supply patient data to LLMs for some task, how do we mitigate such bias? [ties into Question 1 about LLM routers making decisions]

References

Demystifying AI Agents in Clinical Healthcare and Hospital Operations, Paul J. Swider (https://x.com/pswider/status/1896232968693157976)