How clinical agents get built and trained

TL:DR

Everyone is talking about AI agents in healthcare. Today, we’ll talk about what agents are and how they compare to types of automation in the past. There are lots of different ways to train, test, and deploy these agents.

We’ll also cover Amigo, a company that enables providers to build their own agents. We’ll go through the product itself and how they create provider-specific simulated environments to customize and battle test agents. 

The company is taking the approach of building a platform to create multiple agents together vs. selling pre-made agents. We’ll talk about the pros and cons of their approach, and the tradeoffs they’re making.

‍This is a sponsored post.  You can read more about my rules/thoughts on sponsored posts here. 

—

Company Name - Amigo

‍

Amigo has built a platform to build AI agents specific to your practice/clinic. We’ll talk about how they do that in a second.

I got a 2 on AP Spanish but even I know “Amigo” means “friend”. This would be like me starting an AI company and calling it “The Boyz”. 

What is an Agent?

We’ve had automation in healthcare for quite a long time. It might be helpful to think about an agent compared to other types of automation you might have heard about in the past.

Let’s start with Robotic Process Automation. Companies map out workflows a human does - things like clicking on screens, copying and pasting data, etc- and a computer repeats that workflow exactly. This automation exercises no judgment, has a very specific “track” to follow, and frequently breaks if small changes happen in that flow (e.g. a button changes place). It’s brittle and also very narrow in what it can do.

Then we had chatbots. If AI did that “post something from 2016 trend” they’d post a chatbot. Chatbots are a bit more evolved in that they have branching decision trees in the background, so they could handle a wider variety of inputs. However, as soon as they deviated from the path, they broke and required escalation. Plus, there was no chat history so each conversation was brand new, like Memento for chat.

There’s nothing more fun than going to your favorite telecom carrier, being told to use the chatbot with preloaded options, and then having it fail 90% of the time so you need to talk to someone anyway.

Today, we come to agents. Agents are a type of automation that uses this newfangled AI to power it, but they have a few components that make them extra powerful.

1. Differing Personalities - You have the ability to give different agents different  personalities and role types. This allows you to embed different rules, aggressiveness, tone, etc. into a given automation. This matters because it will shape how the underlying language model interprets info, responds, and escalates.
2. Context Graphs - Context graphs give information on HOW decisions are made and the context needed to make those decisions. For example, when a patient asks about their rash the agent needs context about the rest of the patient's condition, what questions to ask in that situation, and what escalation should be, based on some general guidance (with flexibility to adapt). When a doctor sees that message, a similar mental framework is kicking in, developed from seeing that scenario many times and knowing the rest of the patient’s health history.
3. Actions - The agent is allowed to interact and push/pull data to different systems based on the permissions it has. The breadth of what they can do tends to be larger as well, because they can go through API endpoints or use browsers just like how you would use a computer. Today, you even see companies that have specific pathways for agents to interact with their system (e.g. Model Context Protocol).
4. Memory - Agents have memory. This is not only remembering previous times you’ve interacted with them but also how that agent has interacted with different systems and users that might be relevant. Because of the vast amounts of data that could count as memory, an entire hierarchy has to be built that figures out which memory is important for a given task or context. Companies tend to build with very strong opinions about what memory is relevant to pull at which times.

The result is that agents have opinions on how to get things done and flexibility to attempt other things if things don’t go correctly. They have dynamic behavior; if something changes from the given path, they can change with it. When done correctly, they basically function like employees at your job.

What Pain Point is Being Solved? What Does Amigo Do?

What’s hard about deploying agents into clinics is that every clinic is extremely different. They have different rules, see different kinds of patients, have different points of view on triaging, etc. So to learn context and everything that makes agents powerful is really hard.

This is especially true when it comes to patient facing agents or agents that need to handle any level of clinical task. The prior authorization flow will look similar between a rural surgery center and a metropolitan gastroenterologist, but the clinical and patient facing tasks will look very different. The risks are higher and need to be more specific to the setting it’s deployed in, so the agents need to be higher fidelity.

Amigo has a few components to do this.

The first is onsite deployment + an agent factory. They send people to ingest a ton of data from a practice EHR, practice management system, scribes, standard operating procedures, and interviews with the practice managers. By doing this they:

1. Build profiles of the different types of patients a practice sees.
2. Learn the rules of interacting with patients. Both the explicit rules that are written down, and the implicit rules that come from interactions.
3. Learn the personality types to give different agents, what actions to do in different scenarios, and what memory to pull, when.

As you can imagine, this is a massive amount of data in itself, so Amigo has a ton of their own agents that do the work of ingesting a lot of this raw data and creating that foundation. 

The next step is to take the agents that were built and put them in a simulated world with millions of potential patient encounters. It’s like when you’re thinking about a fake argument in your head in the shower and the cool things you’ll say to win, but for AI. Those agents are then judged by another AI that objectively tells them what went wrong, self-analyzes, and makes improvements. This loop helps the agent improve and figure out cases where it won’t be as strong (and therefore may need to escalate to a human in the loop).

With this, they now have agents that are actually usable in patient facing encounters. You can build different kinds of agents for different tasks on top of this foundation, and then they have the tools to monitor how the agent is performing.

–

The Amigo CEO called this the “AI residency program”. AI can pass the MD licensing exams with flying colors, but the doctors learn the real-world implementation when they become residents at a hospital or clinic. The only way that happens is by seeing things a million times and learning from mistakes, which an agent does in your practice-specific simulated environment.

And like real residents, you don’t need to pay them much…too real?

An Agent Example - Side Effect Management

A common use case that gets built with these clinical agents is side effect management.

Let's say you're a digital health company prescribing GLP-1s for weight loss. You’re very original. Patients are texting constantly about side effects. They’re nauseous, the injection site has weird reactions, they didn’t use the pen properly, etc.. They want answers at odd hours and it’s not really worth the time for the practice to have someone clinical answer questions about stool consistency or “you up? wyd”.

Amigo works with the clinical team to build the agent's brain. They define:

• The personality - Warm but direct, avoids medical jargon. 
• The context graph - States for symptom assessment, severity determination, checking medication history, providing guidance, and escalating. 
• The behaviors that override - For example, "if a patient mentions chest pain or severe dehydration, escalate immediately, regardless of where you are in the conversation.”
• What memory dimensions matter here - medications, allergies, how long they've been on the GLP-1, past side effects they’ve had, past dosages, etc. 

The clinical team sets the success metrics: empathy needs to be above an 8, clinical accuracy higher, and appropriate escalation has to be 100% of cases that need it.

Amigo creates thousands of synthetic patients that mirror the practice's actual population and runs the agent through scenarios: "Patient reports mild nausea after first injection"... "Patient has been vomiting for 3 days and can't keep water down"... "Patient is talking about proteinmaxxing but I have no idea what they’re talking about."

Then the AI Judge evaluates each conversation. The transcript, internal reasoning for the agent, and the outcome. Did it check how long the patient has been on the medication? Did the right dynamic behavior fire when the patient mentioned they couldn't keep fluids down? An Amigo engineer reviews, approves, runs it again. This loop continues until it hits thresholds across all metrics.

The agent then gets deployed. Patients can ask questions about side effects they’re feeling and get answers, a new appointment, or get switched to a new drug (with a doc reviewing, async). The clinical team gets dashboards showing where the agent is strong and where it's struggling. The founders can tell their investors they’re AI-enabled care delivery. New edge cases from real patients feed back into the simulation environment.

This process is done for every agent type, clinic, etc.

What Is The Business Model And Who Is The End User?

Amigo charges a base platform fee plus usage-based pricing. Customers get:

• Access to the platform (agent creation, simulation/testing environment, monitoring)
• A forward-deployed "Agent Engineer" embedded with their team. Oh god there are subspecialties of forward deployed engineers now.
• Compute
• A friend

They have two groups of customers. The first are digital health companies doing care delivery. Customer engagement increases and the patients that engage with the patients tend to have higher lifetime values.

The second is traditional providers. Traditional clinics are using these agents to handle patient-facing clinical tasks that might be bottlenecked by clinical hiring (e.g. nurse phone lines, triaging, chasing down care gaps, etc.). This helps practices scale without needing to hire more. 

They work across a range of specialties including primary care, orthopedics, behavioral health, women's health, and wellness/longevity. 

Job Openings

Amigo is looking to triple their team in 2026. They're hiring for:

- Agent Deployment Strategist

- Staff Software Engineer (Infra)

- Staff Software Engineer (Backend)

- Staff Full-Stack Engineer

- Account Executive

- Agent Engineer

You can see all the jobs they’re hiring for here: https://www.amigo.ai/careers

Out-Of-Pocket Take

A few things I think are interesting about Amigo:

Patient-Facing Agents are Here - It’s clear that we’re entering the takeoff phase for patient facing agents. Amazon launched their AI in One Medical, and several startups like Doctronic and Lotus aim to deliver care directly with agents. Federal tailwinds seem to be pushing for more care delivered via agents - ARPA-H has a whole program for agents to help with cardiovascular issues. 

This does feel like the exact right time to have a business around helping providers build and launch their own agentic workflows. Patients are going to eventually start demanding certain workflows be automated as they interact with other services. The stakes are also much higher in these interactions so regulators are going to demand traceability and accountability. Amigo should have the pieces to ride these trends if they execute well.

Building In-House Evaluation + Self Correction Processes - One of the tricky parts of deploying AI is how to evaluate it and then how to correct it (which we’ve talked about in the past). Benchmarks published publicly are useful, but can change dramatically when rolled out to a specific setting.

It’s clear that as AI tools start having interactions at scale, it’s going to be impossible for humans to investigate all of them. Using synthetic environments and having other AI as the judge/oversight is going to be a necessity (with still some humans in the loop to agree on the changes). 

Being able to assess the issue and correct the agent is way easier if you control all of the pieces under the hood and understand how each works. If you’re coordinating several different vendors for a task, it’s going to be impossible to understand what went wrong with the other vendors and what fixes should be made.

The result is hopefully creating better site-specific and interaction-specific benchmarks to see how an agent would perform in the setting they’ll actually be deployed in.

The Benefits of Being Global - Amigo’s first customer was in Australia, and roughly half of their customers are outside of the US. 

I think there’s something interesting about healthcare agents in international markets. 

• Developing countries outside of the US have a lot more cash-pay volume, so increasing the throughput of interactions with agents has a clear return on investment. 
• WhatsApp/natural chat interfaces are used way more vs. in the US where you need to go through a patient portal. Those chat interfaces can make healthcare agent interactions more accessible.
• Because of how bad accessibility is in many countries, they’re more flexible about deployment of AI tools. There are pros/cons around patient privacy and oversight, but this is happening regardless.
• What if agents trained on Johns Hopkins protocols/context graphs can be sold to clinics in other countries? This sort of happens already when Mayo Clinic opens up a hospital in a new country and needs to educate all the local doctors on how they do things; maybe you’ll see something similar to agents.

I just hope when agents study abroad they aren’t half as annoying about it as the people who do.

—

As with any company, Amigo may face challenges, and here are some I’d expect as it grows.

Point Solution vs. Platform - A company that picks one specific use case like prior authorization or appointment scheduling can deploy more quickly and prove ROI faster. Amigo requires more upfront work to actually get set up and requires more involvement from the provider to actually build out the agents. Deployment is on average six weeks, but this also buys time as ROI for clinical use cases can take a while. 

It’s a bit of a race: can the point solution companies targeting lower-risk tasks get in through the door and then get to the more complex clinical stuff? Or, do you need to come in with the point of view that all agents need to be built on top of the platform in order to do complex clinical tasks? Amigo is betting on the latter.

Will Healthcare Orgs Ever Trust Patient-Facing AI? Some organizations just aren't ready to let AI talk directly to patients. They’re cool with the back-office, but there’s way more liability when it comes to the front. You think doctors are going to listen to all this mumbo jumbo about world building and simulations that make the agents safe? They’re worried that if something fails, it’s their ass on the line.

Amigo is making the bet that the trend is towards more practices willing to actually use agents in patient-facing tasks. 

Regulatory Uncertainty - There's ongoing discussion about whether patient-facing AI doing clinical tasks might end up in the software-as-a-medical-device (SaMD) category. The FDA hasn't been super clear here, but in general, this administration seems to be focused on getting AI into clinical practice as quickly as possible.

The Doctronic pilot in Utah allowing autonomous prescribing is a litmus test. If this goes well, we’ll probably see more states pushing to allow this kind of AI agent usage. It’ll also put pressure on doctor’s offices to enable this kind of capability. 

It’s possible, however, that the FDA decides any clinical, patient-facing AI is considered a medical device, which would significantly increase the burden of proof for agents.

Competition - I say this for basically any AI company at this point, but there’s always the question of “what if the EMRs just build this?”. It’s healthcare’s version of “what if Google just copies this?”. 

This is always an existential risk. Amigo’s product seems to require stitching together enough other systems that this would be hard for one single system of record to do. On top of that, it’s more likely that the EHRs realize they can actually make money by charging the agents to interact with their systems and enabling that instead.

The other vector of competition is foundation models getting into this themselves and helping customers directly deploy agents (e.g. OpenAI acquiring the company that made OpenClaw). Will vertical specific training for agents be a huge differentiator? That’s the question for the future.

Conclusion and Parting Thoughts

I’ve been building agents a bit myself. They’re extremely powerful, but even for my very simple agents I have to spend a lot of time giving them instructions and rules.

I do think we’re going to have more agents in healthcare, but it’s clear that we need a full process to create, shape, and monitor them. Amigo’s bet is that they’ve built a system to do that. 

But the real agents? The amigos we made along the way.

Thinkboi out,

Nikhil aka. " Agent Factory? More like Casamigos" 

Twitter: ​@nikillinit​

IG: ​@outofpockethealth​

Other posts: ​outofpocket.health/posts​

‎If you’re enjoying the newsletter, do me a solid and shoot this over to a friend or healthcare slack channel and tell them to sign up. The line between unemployment and founder of a startup is traction and whether your parents believe you have a job.