We are building reinforcement-learning agents to run US healthcare. They start in post-acute care, inside the buildings where the work actually happens.
01 / Intake
They start at the front door, watching every admission a facility makes: who was referred, what the packet said, who got a bed and who did not. Nobody hands them a policy manual. They learn this building's rules from this building's decisions.
What the environment returns
Every admission
observed with the decision a human made, and what happened to that resident afterwards
They start at the front door, watching every admission a facility makes: who was referred, what the packet said, who got a bed and who did not. Nobody hands them a policy manual. They learn this building's rules from this building's decisions.
Every admission — observed with the decision a human made, and what happened to that resident afterwards
Payer rules are not published anywhere. They are learned from what gets approved and what gets denied. The agent assembles the packet a specific payer will say yes to, submits it the moment a referral is accepted, and tracks it from there.
Approved or denied — a clean reward on every submission, which is how payer policy gets learned without anyone publishing it
Continued-stay reviews arrive on the payer's schedule, not the building's, and a missed through-date is a day of care delivered for free. The agent tracks every one of them and has the next review assembled before the days run out.
Days authorized — a countable outcome on every resident every day, so the policy is scored continuously rather than at the end
Orders, assessments, therapy minutes and progress notes become one plan instead of five systems. When something changes on the floor, the plan changes with it — and so does the payer story and the documentation behind it. This is where an agent stops moving paperwork and starts holding clinical context.
Plan against outcome — what the plan predicted, measured against what the resident actually did — the error is the lesson
Eligibility was verified at intake. Authorization is on file. PDPM coding is supported by documentation that already exists. Denials get prevented four steps upstream instead of appealed downstream — which is the difference between a billing tool and an operating system.
Paid or denied — the terminal reward, settled in dollars within weeks — which is why this environment trains and most in medicine do not
Every surface above is a reinforcement environment. Agents watch real decisions inside a real building, learn that building's policy, and then take the work over — with the people who taught them still in the room. What one agent learns, all of them know.
Policy deployed — the agent stops proposing and starts acting, in 100+ live environments today
The simulation
Reinforcement learning needs an agent to act and then be told whether the action was right. Across most of medicine that is either unethical or the answer arrives years later, which is why RL went to games and robotics while healthcare got chatbots.
Post-acute administration is the exception, and it is a narrow one. The decisions are administrative rather than clinical. They repeat dozens of times a day in every building. And the outcome is unambiguous and it is fast: the bed filled or it did not, the authorization held or it lapsed, the claim paid or it was denied.
That combination — dense decisions, clean labels, short feedback loops — is what makes an environment buildable at all. It is a genuinely rare property, and it is the reason this works here first.
A building, modelled
Its queue, its payer mix, its admission criteria, its through-dates, its documentation. Not a generic facility — that one. The agent acts against the model, not against residents.
One decision, scored
A referral judged, a packet assembled, a review filed. The agent proposes, a human decides, and the gap between the two is the gradient. Accepted, corrected or overruled — all three are signal.
What it has learned to do
Once the agent matches the team on a task, it takes that task over and the team supervises. What it learned transfers to every other building's agent without retraining.
No resident is an experiment. The agent’s first thousand decisions are proposals, and the people it is learning from are the ones deciding.
Why now
In 2022 there were fifteen working-age adults for every American over eighty. By 2045 there will be seven. The demand for post-acute care roughly doubles against the labour available to deliver it, and no amount of hiring closes that gap.
There is no version of the future where a forty-minute referral review scales to meet it. The work has to be done by something that learns the building and then does the job — which is why this is an agent problem and not a software problem.
Working-age adults per person aged 80+
Proof
Which matters here for one reason: every building that switches is another live environment. The agents do not get better because we raised more money. They get better because more real decisions are being made in front of them.
“We chose Magicare over every competitor — including Exacare with $37M raised and a 5-year incumbent. Within 90 days, all our homes were full.”
Mary Jean Cappabianca
Corporate Admissions & Social Services Director, Inspira Health Group
“Magicare increased my facility's revenue by $1M.”
Jonah Weiss
Owner, Tacoma Nursing and Rehab
“With Magicare we are never alone. They are true partners.”
Anaili Rendon
Chief Marketing Officer, Excelsior Care Group
The agents
Each agent is trained inside one facility’s real workflow, on that facility’s real decisions. They do not arrive knowing your building. They learn it, and the people who taught them stay in the loop.
INTAKE
Reads every inbound packet, verifies eligibility and ranks fit against the criteria that building actually admits on.
AUTHORIZATION
Assembles and submits prior auth from the referral you already have, then tracks every through-date for continued stay.
PLAN OF CARE
Keeps orders, assessments, therapy and progress notes in one plan that nursing, therapy and admissions all see.
REVENUE
Carries eligibility, authorization and PDPM support into the claim so denials get prevented instead of appealed.
Next
Once an agent holds a building’s clinical and financial context, the same environment reaches the rest of the operation. Nothing new has to be learned from scratch.
Where this goes
It is the right place to start a learning system: the fastest-growing segment in healthcare, decisions made dozens of times a day in every building, and outcomes scored in dollars within weeks rather than years. An agent gets thousands of labelled judgements a month here that it could not get anywhere else in medicine.
But nobody stays in post-acute. More than a quarter of residents go back to the hospital inside a month, and the hospital and the payer both pay for it. The agent holding the post-acute record is already sitting on the seam between the hospital that discharged the patient, the payer funding the stay, and the building delivering it.
Patients, payers and providers: one source of truth for the care pathway between them.
The volume that makes a learning system possible in the first place.
The resident does not stay. Post-acute is a corridor, not a destination.
Which is money the hospital and the payer are already losing on this seam.
Why it compounds
A competitor cannot start where we are by training a better model.
They would need a hundred buildings willing to let agents watch them make real decisions, for long enough to learn how each one actually operates — while those buildings keep running on something else in the meantime. That is the part that took years, and it is the part that gets harder to catch every month it runs.
It also compounds. What one agent learns, all of them know, so every building that comes online makes the next one cheaper to win and faster to run. No manual retraining, no starting over — which is why the agents behave like staff who have worked in a building for years rather than a model that read a manual.
The team
Agents that make operating decisions inside a regulated building are a systems problem before they are a modelling problem. This team has shipped that kind of system in markets where being wrong is expensive.
CEO
Team leader at industry leading firms; designed and built decision-making tools for managing $100B in AUM at Millennium Management.
CTO
Real-time data and interoperability specialist; built a 50+ person consultancy serving major enterprise customers.
COO
Strategic operations and customer success leader with 8+ years building teams across AI, SaaS, services and healthcare.
CRO
Former top-VC partnerships lead; guided 10+ cyber, fintech and devtool startups from day one to product-market fit, with exits to Palo Alto Networks, IBM and Tipalti.
Founding Engineer
Engineering leader and edge AI researcher; Softmax engineer, Uber hacker and Thiel Fellow.
If you invest in companies at this layer, we should talk.
An agent that runs a building has to be trained inside one. Magicare is already inside 100+ of them, and every decision made there is signal the next agent starts with. That is not a dataset a competitor can license.
HIPAA · SOC 2 · BAA available