Before we go much further, we need to carve out what we mean by "AI agents." People use the term loosely enough that deciphering what AI agents are gets murky.
Sometimes, people dub any software with sensors and strong if-then logic as an "agent." If we label any software with sensors and strong if-then logic an "agent," though, a lot of medical software that doesn't seem terribly "agentic" gets lopped into the "agent" bucket (e.g., if patient A's blood oxygen saturation drops below X%, administer Y liters per minute of supplemental oxygen). Similarly, AI "assistants" like Siri or Alexa, though they enjoy a larger set of possible "actions," still seem too reactive to earn "agent" status—they’re mostly passive, responding narrowly to our specific requests.
A more satisfying, intuitive definition of AI agents might be software-based entities that proactively pursue constrained (though not completely spelled-out) goals. These goals could be triggered by events, assigned by humans, or even other AI agents, leading to hierarchical divisions of labor, with 'managerial' agents orchestrating specialized ones—or even simulating entire networks of people.
At first glance, this might not seem terribly revolutionary. After all, if we view AI agents as functions, their interactions appear similar to traditional programming: functions calling other functions, each taking in data, transforming it, and passing it along to the next function. This raises the question: Are AI agents truly transformative, or are they merely a fundraising-focused rebranding of existing computing approaches?
To answer this, let's examine how we currently interact with computers.
More Than Just Functions
Think about how we currently get computers to do what we want them to do.
We write painstakingly precise instructions, specifying how to decompose some larger task, within the rigid syntax of our chosen language (C, Python, Haskell, etc.), to craft functions that call functions that call functions—and so on.
And no matter how meticulous we are, bugs sneak in, revealing that our how-to instructions weren't exactly as exact as we’d aimed for.
So we clarify, and then all is well and good—for a while.
But as we add new features, more imprecisions emerge; maybe parts of the new code don’t integrate seamlessly with the old, or maybe previously missed something that now bears its teeth.
So, we clarify again, and again, and again, ad nauseam.
This painstaking cycle can require entire development teams and can sometimes cause software projects to balloon in time, effort, and cost, edging intended “automation” projects toward a net negative.
The allure of AI agents is that they might free humans from this annoying cycle of writing explicit instructions and instead (imperfectly) figure out how to achieve goals on their own—by reflecting, deciding, acting, and learning over time—without needing every tiny, intermediate step to be programmed for them. This approach, of course, entails ceding some control and determinism for some convenience and stochasticity, a risky move in healthcare settings where human lives are on the line. We'll explore these risks later, but first, let’s skim through AI agents’ core ingredients and the benefits that they offer.
AI agents typically leverage the following components:
- 📚 Language - via LLMs, speech-to-text models (like Nova-3), and text-to-speech models
- 🧠 Memory - often through retrieval-augmented generation (RAG)
- 📝 Planning - by breaking down complex tasks into manageable steps, often via prompt engineering techniques like Chain-of-Thought, Tree-of-Thought, or Graph-of-Thought
- 🛠️ Tools - typically by calling APIs via function calling
These all work together to empower AI agents with a degree of autonomy. What would this look like in a real-world healthcare setting? We’ll go into a thorough, hypothetical example later, but for now consider the simple task of a nurse creating a referral for a patient to see a different specialist. One AI agent might transcribe that nurse’s verbal request; another correct transcription errors; a third analyze the corrected text to determine which databases or APIs to query for relevant patient data, while a fourth searches for that specialist's contact info and then calls and schedules the appointment. Each agent might be assigned a narrow role, and yet they all collaborate on a broader task. Healthcare constantly requires such coordination because healthcare data is frequently compartmentalized across multiple electronic systems, and treatments are often done across multiple facilities.
Some preliminary research suggests that this kind of "multi-agent" interaction might even be effective beyond real-world applications. In a recent experiment called "Agent Hospital," Li et al. tested if AI agents, acting like patients, nurses, and specialist doctors, could learn to work together in a simulated hospital. Across 10,000+ patient simulations, the AI "doctors" improved at their jobs, even beating GPT-4 (the simulation's base LLM) at diagnosing respiratory diseases in questions from the MedQA benchmark. This hints at something akin to AlphaZero's famed ability to improve at games like Go and chess by repeatedly playing against itself: AI agents also seem to learn more by repeatedly interacting with one another than they do from static datasets alone.
The early results and ideas that we're seeing from AI agents—from coordinating healthcare tasks to learning through simulation—hint at their transformative potential. To fully appreciate what makes them different, we’ll rewind to an earlier generation of medical AI—rule-based "expert systems"—which, despite impressive capabilities in narrow domains, never quite gained widespread clinical adoption. Check out part 3 of this series to dive in!
Final note: If you’d like to code with an AI agent yourself, check out this API! It works pretty well out-of-the-box, in our opinion 😉
To see Part 1 of this series, click here.