ORTHO-X : From AI Assistants to Agentic Orthopaedics

Why the Future of Surgical Intelligence Is Multi-Agent

Artificial intelligence in medicine is moving remarkably quickly.

We already have systems that can interpret images, summarize clinical records, retrieve scientific evidence, draft documentation, analyze surgical video and converse with patients.

Yet there is a fundamental problem.

A patient does not experience healthcare as a collection of isolated AI tasks.

A patient experiences a journey.

For an orthopaedic patient, that journey may begin with an injury or chronic pain, continue through referral, imaging and diagnosis, require a treatment decision and perhaps an operation, and then extend for months through rehabilitation and outcome measurement.

The difficult question is therefore no longer simply:

Can an AI model perform a medical task?

The more important question is becoming:

How do many different AI capabilities work together safely around the same patient?

That is the reason we believe the next phase of Surgical Intelligence will be multi-agent.

And it is why MAIVAN.ai is adapting modern multi-agent design principles to the ORTHO-X architecture.


Medicine is already a multi-agent system

Consider an older patient arriving at hospital with a fractured hip.

Within a relatively short period of time, many different people and systems may become involved:

the emergency department, radiology, orthopaedic surgery, anesthesiology, nursing, physiotherapy, rehabilitation, hospital administration, the implant supplier, the insurer and eventually perhaps a clinical registry.

Each sees a different part of the patient’s case.

Each performs a specialized function.

The challenge of healthcare has never been a lack of specialists.

It has often been a lack of continuous coordination around the patient.

Artificial intelligence risks recreating exactly the same fragmentation digitally.

A hospital may soon have:

  • one AI interpreting radiographs;
  • another generating clinical notes;
  • another retrieving evidence;
  • another assisting surgical planning;
  • another analyzing the operation;
  • another preparing discharge;
  • another contacting the patient at home;
  • another collecting outcome data.

All may individually perform well.

But who coordinates them?

Which agent is allowed to see which information?

Which result becomes part of the clinical record?

What happens when two systems disagree?

When must a surgeon intervene?

Who remembers why a decision was made?

And how do we determine later whether the entire AI-supported pathway actually improved the patient’s outcome?

These questions define the next generation of clinical AI.


From one AI assistant to a digital clinical team

Much of today’s healthcare AI is still presented as an assistant sitting beside a doctor.

That model is useful, but incomplete.

The emerging architecture is closer to a digital clinical team.

Different agents specialize in different tasks.

One interprets imaging.

Another classifies the injury.

Another retrieves relevant evidence.

Another proposes treatment alternatives.

Another watches for contraindications.

Another tracks what happened during surgery.

Another collects postoperative outcomes.

The intelligence does not lie only inside any one agent.

Increasingly, it lies in how these agents are coordinated.

This is the central lesson of modern multi-agent system design.

For ORTHO-X, we translate that lesson into several distinct components.


OrthoFlow: the conductor, not another instrument

At the center is ORTHO-X OrthoFlow™.

Think of an orchestra.

Excellent musicians do not automatically produce an excellent performance simply by sitting in the same room.

Someone must coordinate sequence, timing and interaction.

That is the role of OrthoFlow.

OrthoFlow determines what needs to happen next in a clinical episode, which specialized capability should perform that task, what information it may use, what output is expected and whether human approval is required before the pathway continues.

That distinction is important.

OrthoFlow is not intended to become a giant medical AI that knows and does everything.

Instead, it orchestrates many specialized capabilities.


OrthoSkills: AI with boundaries

We call these specialized capabilities OrthoSkills™.

An OrthoSkill should have a narrowly defined clinical purpose.

For example, a fracture-classification OrthoSkill might receive radiographs and relevant case information and return:

  • a structured AO/OTA classification;
  • the evidence supporting the classification;
  • a confidence score;
  • any uncertainty requiring review.

Importantly, the system should also know what that agent is not authorized to do.

Classification does not automatically imply treatment authority.

Image interpretation does not automatically imply permission to change a medical record.

A recommendation does not automatically become a clinical decision.

For healthcare, the boundary around an AI capability may ultimately be as important as the capability itself.

We therefore define an OrthoSkill not merely by what model it uses, but by its:

purpose, inputs, permitted tools, outputs, confidence, authority, escalation rules, provenance and evaluation evidence.

This moves clinical AI away from vague claims of “intelligence” and toward testable clinical capabilities.


The Case Capsule: one shared version of the case

Multi-agent systems create another problem.

If every agent maintains its own understanding of the patient, inconsistencies can rapidly accumulate.

ORTHO-X therefore anchors the episode to a Case Capsule™.

The Case Capsule is a provenance-aware longitudinal representation of the patient’s orthopaedic journey.

It can connect:

diagnosis, imaging, classifications, clinical decisions, surgical plans, implants, operative events, rehabilitation, complications and patient-reported outcomes.

It can also record what an AI proposed, what evidence it used and what the responsible clinician ultimately decided.

In technical language, it provides shared state.

In clinical language, it means:

Everyone works from the same case.

Instead of allowing agents to conduct endless unstructured conversations with each other, a safer pattern is:

read the authorized case state → perform a bounded task → produce a structured result → validate it → write the provenance-stamped result back.

This makes the Case Capsule something like the transaction ledger of the clinical episode.


Human authority remains explicit

A multi-agent medical system should not be confused with an autonomous hospital.

We see human authority as part of the architecture rather than as an inconvenience added afterwards.

One useful design pattern is:

Actor → Critic → Clinical Authority → Recorder.

Imagine a surgical planning agent.

It proposes a treatment.

A second capability checks for relevant contraindications, inconsistencies or missing information.

The surgeon reviews the combined result and accepts, changes or rejects it.

The Case Capsule records the decision and its provenance.

For lower-risk tasks, less intervention may be necessary.

For higher-risk decisions, independent verification and explicit clinician confirmation may be required.

The important point is that autonomy becomes configurable and evidence-dependent.


Not everything should be an agent

There is considerable excitement around autonomous AI, but healthcare should resist a simple equation:

more autonomy = more progress.

Many clinical processes should remain deterministic.

If a workflow can be executed safely and predictably through ordinary software, there is little reason to ask a language model to improvise.

The ORTHO-X architecture therefore combines four things:

rules + workflows + AI agents + human authority.

Agentic reasoning should be used where it adds value—for interpretation, ambiguity, synthesis, adaptation and complex decision support.

Deterministic workflows should remain deterministic.

This is not a limitation of agentic medicine.

It is good engineering.


A hip-fracture journey as a multi-agent OrthoFlow

Take the example of a proximal femur fracture.

A future OrthoFlow could connect:

clinical intake → imaging → fracture classification → risk assessment → treatment planning → surgeon confirmation → operative planning → surgical recording → implant identification → postoperative assessment → discharge → rehabilitation → patient-reported outcomes.

Different OrthoSkills could operate at each stage.

But all remain connected through the same patient journey and the same Case Capsule.

The result is fundamentally different from today’s collection of disconnected AI applications.

The unit of intelligence becomes not merely the individual prediction.

It becomes the entire episode of care.


Interoperability: agents must work across institutions and vendors

No hospital will obtain every future clinical AI capability from one supplier.

Hospitals already operate heterogeneous technology environments, and this diversity will increase as AI expands.

That means Surgical Intelligence must be designed for interoperability.

ORTHO-X therefore distinguishes between two types of connection.

The first connects an agent to tools and data.

Examples include the hospital EHR, PACS, a clinical database, a medical device, a registry or a scientific knowledge source.

Technologies such as APIs, FHIR, DICOM and increasingly the Model Context Protocol can support these interactions.

The second type connects agents to agents.

A hospital agent may need to interact with a patient agent.

An imaging agent may provide structured information to a planning agent.

A hospital may one day interact with an implant manufacturer’s post-market surveillance agent or a registry’s outcomes agent.

Emerging agent-to-agent protocols are beginning to make such architectures technically practical.

For ORTHO-X, this reinforces an important principle:

We do not need to build every AI capability. ORTHO-X needs to make heterogeneous capabilities usable together safely.


Why this matters for hospitals

Hospitals are increasingly confronted with a difficult strategic question.

Which AI should they buy?

That question may soon become impossible to answer product by product.

Hospitals will instead need an infrastructure for deciding:

  • which agent may participate;
  • which clinical task it may perform;
  • where it may execute;
  • what data it may access;
  • which version is active;
  • how it has been validated;
  • whether human confirmation is required;
  • how its performance is monitored;
  • how it can be replaced without redesigning the entire workflow.

This shifts the problem from AI procurement toward AI governance and orchestration.

That is one of the central purposes of ORTHO-X.


The Case Arena: Test the journey, not only the answer

Clinical AI evaluation today often focuses on isolated questions.

Did the model classify the image correctly?

Did it generate the right diagnosis?

Did it produce an acceptable answer?

These questions are necessary but increasingly insufficient.

In a multi-agent clinical environment we also need to know:

Did the right agent receive the right information?

Did it call the correct tool?

Did it recognize uncertainty?

Did it escalate when appropriate?

Did the surgeon see the critical evidence?

Was the final plan different from the initial AI proposal?

Was that plan actually carried out?

And ultimately:

What happened to the patient?

This is why MAIVAN.ai is developing the Case Arena™.

A clinical case can be replayed through competing agent and workflow configurations.

Different models can be compared.

Different planning strategies can be evaluated.

Human-only workflows can be compared with human-plus-AI workflows.

The planned treatment can later be compared with the actual procedure and the resulting outcome.

Instead of benchmarking isolated AI answers, we can begin benchmarking clinical trajectories.


From Case Arena to Global Collaboration Center

This concept also helps explain the purpose of the proposed ORTHO-X Global Collaboration Center for Surgical Intelligence.

The GCC is not intended to centralize all patient data.

Its primary function is to create an environment in which clinical agents and workflows can be:

designed → simulated → evaluated → validated → governed → deployed → monitored → improved.

Sensitive clinical information can remain within the appropriate hospital or sovereign environment.

What can travel are validated models, standardized OrthoSkills, evaluation protocols, synthetic cases, performance evidence and authorized structured outputs.

The GCC therefore becomes an infrastructure for collaborative clinical engineering rather than merely another healthcare data lake.


The missing agent: the patient

There is another important consequence of thinking about healthcare as a multi-agent system.

Most emerging architectures naturally create agents for institutions.

There will be hospital agents.

Physician agents.

Insurer agents.

MedTech agents.

Registry agents.

But who represents the patient?

We believe the patient should increasingly have an agent of their own.

Through concepts such as MUOVITI.life, that patient-side agent could help maintain the individual’s Case Capsule, explain care choices, manage permissions, collect outcomes, arrange follow-up, support rehabilitation and request second opinions.

The conceptual shift is important:

The patient should not simply remain a data source consumed by everyone else’s agents.

The patient should have a digital representative within the system.

This could become particularly powerful for outcome-based healthcare.

Patient-reported outcomes and longitudinal recovery data are valuable not only to the treating physician but also to quality systems, registries, insurers, procurement organizations and device manufacturers.

If the patient becomes the persistent anchor of that information, patient-centered care acquires a new technical foundation.


Outcomes are the ultimate evaluation layer

Multi-agent engineering introduces sophisticated new metrics.

We can measure latency, tool calls, routing accuracy, agent disagreement and workflow completion.

All matter.

But medicine ultimately gives us a much stronger test.

Did the patient benefit?

For that reason, the ORTHO-X evaluation hierarchy should extend all the way from:

model performance → OrthoSkill performance → workflow performance → clinical process → patient outcome.

This creates the possibility of connecting technical AI evaluation with Value-Based Healthcare.

A clinically elegant multi-agent system that produces no measurable improvement in outcomes, access, efficiency or cost has not completed its job.


From model-centric AI to care-centric AI

The first generation of medical AI was largely model-centric.

We asked:

How accurate is the algorithm?

The emerging generation must become care-centric.

We will increasingly ask:

How well does the complete human-and-AI system care for the patient?

That requires a different architecture.

It requires orchestration.

Shared state.

Clear authority.

Interoperability.

Provenance.

Evaluation.

And outcomes.

This is the direction in which we are developing ORTHO-X.


An open R&D project

To make these ideas concrete, MAIVAN.ai is adapting the open-source work accompanying Victor Dibia’s Designing Multi-Agent Systems into an ORTHO-X multi-agent design environment.

The upstream project is valuable because it approaches agents from first principles and explores both deterministic workflows and more autonomous forms of orchestration, together with human intervention, observability and evaluation.

The ORTHO-X adaptation will progressively add the requirements that become essential when those patterns enter medicine:

clinical authority, Case Capsules, OrthoSkills, healthcare interoperability, sovereign deployment, patient agency, clinical simulation and outcome-based evaluation.

The intention is not merely to discuss agentic healthcare.

It is to make the architecture testable.


One patient. Many agents. One accountable journey.

Artificial intelligence will almost certainly become more distributed.

Hospitals will use more models, not fewer.

Medical-device companies will develop their own intelligent systems.

Patients will use personal health agents.

Professional societies will encode clinical knowledge.

Registries will increasingly become computational.

Robots and navigation systems will become intelligent participants in the operating room.

The challenge will no longer be scarcity of artificial intelligence.

It may become an abundance of intelligence without coordination.

That is the problem ORTHO-X seeks to address.

Our working proposition is simple:

One patient. One Case Capsule. Many specialized agents. Human authority. Measurable outcomes.

If that architecture can be made safe, interoperable and clinically accountable, multi-agent systems may become much more than another development in artificial intelligence.

They may provide the technical foundation for truly Patient-centered Surgical Intelligence.

Inspect the open code repository at
github.com/MAIVAN-ai/ORTHO-X-Agentic-Architecture

ORTHO-X : Stewarded by the ERMOS Foundation