Executive Overview

MetaBrain's technology is built around a patented, sensor-informed, closed-loop AI architecture that enables intelligent agents to engage with people, gather first-person context, interpret real-world conditions, and adapt their behavior to determine appropriate next actions.

The fundamental principle is that sensor measurements alone do not provide sufficient context for intelligent decision-making.

A sensor can measure a glucose level, heart rate, EEG pattern, blood pressure, respiratory change, movement, vocal characteristics, or other physiological, behavioral, or environmental condition. However, the measurement alone cannot necessarily explain what the person is experiencing, what caused the condition, or what action is appropriate.

MetaBrain addresses this limitation by combining real-time sensor measurements, first-person human interaction, retained historical context, and AI-driven evaluation within an iterative closed-loop process.

An agent engages the individual, gathers information that sensors cannot independently provide, evaluates the combined information against relevant baselines, session memory, historical observations, and applicable rules, and determines its next question, intervention, or authorized action. As new information becomes available, the agent reassesses the situation and adapts its behavior.

The process repeats until the assigned objective is completed, another action is required, or the interaction is appropriately escalated or terminated.

Importantly, the interaction does not have to begin with a sensor event. An agent may be initiated by a scheduled task, a user request, a clinician instruction, a workflow requirement, a sensor event, a change in sensed state, or another defined condition. Sensor measurements and human context are then combined as part of the ongoing decision-making process.

A Common Closed-Loop Architecture Across Specialized AI Agents

MetaBrain extends this underlying intelligence through a specialized, rules-governed, multi-agent orchestration architecture.

Rather than relying on a single general-purpose AI agent to perform every task, recognize every problem, investigate every condition, and determine every outcome, the architecture distributes responsibility among specialized agents operating within defined authority boundaries.

Task Agents, Intervention Agents, and Resolution Agents share the same fundamental adaptive closed-loop architecture.

Each agent has an assigned objective and can repeatedly:

1. Obtain relevant sensor measurements and other available observations.
2. Engage the human subject to gather first-person contextual information.
3. Combine and evaluate the information using AI, retained context, historical baselines, and applicable rules.
4. Determine the next question, interaction, intervention, or authorized action.
5. Reevaluate the results and adapt subsequent behavior as additional information becomes available.

The agents differ primarily in their assigned objectives, specialized instructions, permissions, and completion criteria, rather than requiring fundamentally different underlying technologies.

This means the intelligence is not confined to the original Task Agent. Intervention Agents and Resolution Agents also conduct adaptive, context-gathering, sensor-informed interactions to accomplish their respective objectives.

The architecture supports multiple agents within each category, allowing individual agents to be configured for specific tasks, conditions, populations, protocols, or operational requirements.

Task Agents: Performing the Primary Objective

Task Agents perform the work for which an interaction was initiated.

Examples include clinical assessments, electronic clinical outcome assessments (eCOA), patient-reported outcomes (ePRO), physiological monitoring, medication adherence, neurological monitoring, behavioral assessments, device-guided workflows, and other healthcare or operational activities.

A Task Agent follows its assigned objective while continuously or periodically gathering relevant information, engaging the person, evaluating sensor measurements and human responses, and adapting its next interaction or action.

For example, during a clinical assessment, a Task Agent may obtain a physiological measurement, ask the participant what they are experiencing, evaluate the combined information, and determine whether to continue the assessment, ask a follow-up question, obtain another measurement, or take another permitted action.

The Task Agent maintains an execution state identifying its current objective, completed steps, pending questions, sensor information, human responses, applicable rules, and other information required to continue the task.

This state can be preserved when another agent needs to intervene, allowing the original task to resume without restarting or requiring the person to repeat previously provided information.

Intervention Agents: Intelligent Observation, Investigation, and Adaptive Engagement

Intervention Agents are a supervisory intelligence layer that observes the behavior, interactions, and operational states of both Task Agents and Resolution Agents.

Rather than relying on a general-purpose Task Agent to recognize and address every possible exception, MetaBrain can deploy multiple specialized Intervention Agents, each assigned to monitor one or more defined conditions.

Examples include:

• Suicidal ideation or other safety concerns.
• Confusion or misunderstanding.
• Disengagement or non-responsiveness.
• Inconsistent, conflicting, or incomplete information.
• Missing or invalid sensor measurements.
• Unexpected physiological or behavioral changes.
• Inappropriate language or emotional distress.
• Device problems, protocol deviations, or quality concerns.

Each Intervention Agent operates according to a defined objective, relevant monitoring criteria, specialized conversational instructions, permitted actions, escalation requirements, and applicable guardrails.

An Intervention Agent may observe conversational exchanges, human responses, sensor measurements already collected by the Task Agent, derived behavioral information, structured agent states, and outcomes of actions taken by another agent. The Intervention Agent does not independently gather real-time sensor measurements. It works with the information already available from the active session. If sensor data is missing or insufficient, obtaining it is the responsibility of the Task Agent, not the Intervention Agent.

Observation may be continuous, periodic, or responsive to defined conditions. An Intervention Agent does not need to interrupt an interaction merely because it is observing it.

When a condition requiring attention is detected, the Intervention Agent can initiate an adaptive intervention without requiring the person to recognize the problem, select another agent, or request assistance.

Adaptive Intervention Behavior

An Intervention Agent does considerably more than identify a problem or route the conversation elsewhere.

It engages, investigates, evaluates, and adapts.

The agent can temporarily assume conversational control, ask questions specific to the detected condition, gather first-person context through its adaptive questioning, and evaluate the combined information using the data already available from the session. Unlike Task Agents and Resolution Agents, which gather and use real-time sensor data to determine what to do, the Intervention Agent's role is strictly related to the detected problem. It determines whether what it has observed constitutes an issue, whether continuing with the Task Agent is appropriate, and if not, hands off to a Resolution Agent that may return the participant to the Task Agent, escalate to human staff, or end the session.

Its questions and actions are not limited to a fixed script. The agent can adapt the content, sequence, timing, and nature of subsequent interactions according to what it learns.

For example, an Intervention Agent may initially ask a clarification question, evaluate the response together with the sensor information already available from the Task Agent's session, identify a remaining uncertainty, and ask a different follow-up question. It may repeat this process until it has sufficient information to assess the condition or determine that additional action is necessary.

The agent may compare newly acquired information with previous responses, baseline measurements, historical patterns, prior interventions, or expected outcomes.

This allows the Intervention Agent to assess not only whether a condition was detected, but also why it may have occurred, whether it remains present, whether the intervention was effective, and what should happen next.

The Intervention Agent can resolve some conditions through its own adaptive interaction. Other conditions may require a specialized Resolution Agent, human involvement, or an external workflow.

Importantly, the Intervention Agent also observes Resolution Agents. If a new problem arises during a resolution process, it can detect that condition and initiate another adaptive intervention while the system preserves the Resolution Agent's execution state.

Resolution Agents: Specialized Corrective Action and Outcome Evaluation

Resolution Agents are responsible for addressing conditions that require corrective action beyond the scope of an initial intervention.

Rather than sending every unresolved problem to one general-purpose agent, MetaBrain can deploy specialized Resolution Agents for particular problem classes.

Examples include device connectivity, invalid measurements, persistent confusion, incomplete data, protocol deviations, behavioral concerns, and safety-related conditions.

Each Resolution Agent operates through the same adaptive closed-loop architecture as the Task and Intervention Agents.

It receives the relevant context already gathered, including the original task, detected condition, human responses, sensor observations, intervention findings, and unresolved information.

The Resolution Agent then pursues its assigned corrective objective, gathering additional sensor measurements and first-person context, evaluating the results, determining permitted corrective actions, and reassessing whether the intended outcome has been achieved.

Depending on the condition and its authority, a Resolution Agent may:

• Investigate and correct a technical or operational problem.
• Obtain missing or replacement measurements.
• Conduct additional adaptive questioning.
• Guide the person through an authorized corrective procedure.
• Evaluate whether the corrective action was successful.
• Return the participant to the Task Agent to resume from its preserved state.
• End the session as a resolution when the condition is resolved and no further task work is needed.
• Escalate to a clinician or human staff when the condition requires qualified human involvement.
• Recommend continued resolution when additional corrective work is needed.

Resolution is not simply the execution of a corrective instruction. It is an adaptive process that evaluates whether the intended result was actually achieved.

The Intervention Agent may continue observing the Resolution Agent throughout this process, identifying further conditions requiring investigation or corrective action.

Rules-Governed Orchestration and Decision Authority

MetaBrain separates adaptive conversational intelligence from authority over workflow transitions and critical decisions.

A dedicated orchestration controller governs which agent is permitted to act, which agent controls the human-facing conversation, when an intervention or resolution may begin, and whether an interrupted activity may resume.

The controller evaluates structured information produced by the agents against stored rules, applicable protocols, safety requirements, and authorization criteria.

An Intervention Agent may determine that a condition appears resolved and produce a structured intervention state. A Resolution Agent may determine that its corrective objective has been achieved and produce a structured resolution state.

These states can identify the detected condition, supporting information, investigation findings, remaining requirements, resolution status, recommended disposition, and other relevant information.

The orchestration controller then determines which transition is authorized.

Possible outcomes include continued task execution, further intervention, invocation of a specialized Resolution Agent, escalation to a human or external workflow, or termination.

This separation allows AI agents to engage in flexible, intelligent, context-sensitive interactions without giving any individual conversational agent unrestricted authority over safety-critical or workflow-critical decisions.

Rules-based quality-assurance functions can independently validate completeness, consistency, sensor quality, protocol compliance, safety requirements, and other conditions before transitions are authorized.

Concurrent Monitoring, Priority Arbitration, and Preemption

Multiple Intervention Agents may observe the same ongoing interaction simultaneously, each monitoring its assigned condition.

More than one condition may arise during a Task Agent's interaction or a Resolution Agent's corrective process.

MetaBrain's orchestration architecture includes priority arbitration and preemption capabilities to manage competing conditions without allowing agents to issue conflicting instructions.

When multiple conditions are detected, the controller evaluates their priority, urgency, severity, safety implications, protocol requirements, and permitted interruption rules.

The controller may authorize one Intervention Agent to assume conversational control while retaining other detected conditions as pending.

Pending conditions preserve their supporting evidence, priority, current status, and relevant context.

If a higher-priority condition emerges, the controller may interrupt an active resolution, preserve its execution state, and authorize another agent to engage.

After an intervention or resolution concludes, the controller reevaluates pending conditions using the most recent sensor measurements, human responses, validated findings, and interaction history.

A previously detected condition may no longer require attention because information gathered during another intervention has already resolved it.

Alternatively, the condition may remain active and require its own specialized intervention.

Only one agent needs to hold primary conversational control at a time, even while multiple specialized agents continue observing and evaluating conditions in the background.

Priority rules can be configured for different clinical protocols, populations, applications, and operational environments. Safety requirements may override ordinary scheduling priorities.

The architecture records relevant intervention decisions, deferrals, preemptions, handoffs, and outcomes to support traceability and auditability.

Longitudinal Context, Memory, and Continuity

A central capability of the architecture is the preservation and reuse of information across agents, interventions, resolutions, and successive interactions.

The system can retain sensor measurements, baseline values, prior human responses, conversational history, previous interventions, corrective actions, outcomes, and other relevant contextual information.

This information is not limited to the immediate conversation. Authorized longitudinal records may be used across sessions to inform subsequent decisions.

Each agent can receive the information relevant to its assigned objective without requiring the individual to repeat information that has already been collected.

For example, an Intervention Agent investigating an unexpected physiological measurement can evaluate the current sensor reading alongside the person's explanation, prior measurements, previous explanations, and known baseline.

A Resolution Agent subsequently invoked to address the condition can receive that same context, together with the findings from the intervention.

When the issue has been appropriately addressed, the original Task Agent can resume with its prior execution state intact and relevant new findings incorporated into the continuing interaction.

The system therefore maintains continuity not merely of the conversation, but of the accumulated knowledge, decisions, actions, and outcomes that inform subsequent agent behavior.

Illustrative Example: Adaptive Behavioral Intervention

Consider a Task Agent conducting an ongoing assessment or structured interaction.

During the conversation, the person begins using inappropriate language.

A specialized Intervention Agent observing the interaction identifies the behavior as a condition requiring attention.

The condition is recognized through conversational content; no sensor-triggered event is required.

The orchestration controller preserves the Task Agent's current position, previously collected information, and pending work.

The Intervention Agent's role is limited and focused. It does not conduct an extended investigation or corrective process itself. It validates what it believes it observed and asks the person to confirm.

For example, it may say: "I noticed you seem frustrated. Is that true, and if so, what would you like to do about it? Would you like to take a break and return later, or continue?"

The person's response determines what happens next.

If the person says the observation was not accurate, the Intervention Agent determines the condition does not require further action and the controller returns the person to the Task Agent at the exact point where it was interrupted.

If the person confirms the observation and indicates they would like to return to the task, the Intervention Agent determines the condition is addressed and the controller resumes the Task Agent from its preserved state.

If the person confirms the observation but does not wish to return to the task, the Intervention Agent hands off to a specialized Resolution Agent to work on resolving the underlying issue.

If the person indicates they are done and do not wish to continue, the Intervention Agent can end the session.

The person experiences a single continuing conversation.

Behind the scenes, the system has performed supervisory observation, condition detection, preservation of task state, validation of the observed condition, a decision based on the person's response, and rules-governed return, handoff, or termination.

A separate Resolution Agent was not required because the Intervention Agent was able to validate the condition and determine the appropriate outcome within its limited authority.

The same architectural principles apply when the condition arises during a Resolution Agent's activity or when the intervention requires additional specialized corrective action.

A Seamless, Human-Centered Experience

Although the internal architecture may involve multiple specialized AI agents, monitoring functions, decision-control components, and quality-assurance processes, the person does not need to manage that complexity.

The individual can remain within the same conversation, voice interface, avatar, application, or other interaction channel.

The system internally determines which specialized agent should engage, preserves the relevant context, and manages transitions without requiring the person to select agents, restart tasks, or repeatedly explain the same situation.

This is particularly important in healthcare, clinical research, remote monitoring, and other settings where participants may be ill, confused, distressed, distracted, or otherwise unable to manage complex technical workflows.

The technology adapts around the person rather than requiring the person to adapt to the technology.

The MetaBrain Architectural Distinction

MetaBrain's approach is not simply a chatbot connected to sensors, nor a collection of specialized agents passing conversations between one another.

It is an integrated architecture combining:

• Adaptive, sensor-informed closed-loop intelligence operating across Task, Intervention, and Resolution Agents.
• First-person contextual inquiry that adds information unavailable from sensor measurements alone.
• Supervisory Intervention Agents that observe both Task and Resolution Agents and independently investigate conditions requiring attention.
• Adaptive behavioral and operational intervention in which subsequent questions and actions change according to newly acquired information.
• Specialized corrective Resolution Agents that pursue and evaluate defined outcomes.
• Longitudinal memory and shared contextual information supporting continuity and increasingly informed decisions across interactions.
• Rules-governed orchestration and decision authority controlling transitions, escalation, resumption, and permitted actions.
• Concurrent monitoring, priority arbitration, and preemption managing multiple conditions without conflicting conversational control.
• Structured intervention and resolution states, independent quality assurance, and auditability supporting controlled, traceable workflows.
• Seamless human-facing continuity that preserves the original task and minimizes unnecessary repetition or fragmentation.

The architecture combines the adaptability of AI with defined operational boundaries and a continuous flow of sensor measurements, human context, evaluated outcomes, and retained knowledge.

Its central distinction is that every specialized agent can intelligently gather context, evaluate changing real-world conditions, and adapt its behavior within a common closed-loop process, while supervisory intervention and rules-governed orchestration determine how those agents cooperate, resolve problems, and continue the human interaction.

This provides a foundation for deploying increasingly capable agentic AI across healthcare, clinical research, wearable technologies, and other sensor-informed environments while preserving specialization, controlled decision authority, and human-centered interaction.

The Clinical Opportunity

Clinical trials are becoming continuous, sensor-rich, and increasingly always-on.

The volume of signals is expanding rapidly while site bandwidth, participant attention, and data quality remain constrained by human workflows.

The opportunity is a safety-first agent layer that operates between participants, sensors, and trial systems—collecting better data, reducing operational burden, and helping mediate risk in real time, 24/7.

Organizations can adopt generic platforms and work within their constraints, or they can establish proprietary agent capabilities designed around their own protocols, populations, endpoints, workflows, and risk models—creating an institutional capability rather than another vendor dependency.

MetaBrain enables the second approach.

We transfer the architecture, knowledge, patterns, decision frameworks, orchestration methods, and operating practices required for your organization to create and own its implementation within its existing technology environment.

Your internal teams and technology partners build and operate the system.

MetaBrain provides the architecture, knowledge transfer, guidance, training, architecture-conformance review, and patent license required to establish the capability correctly and expand it over time.

Safety by Architecture

One critical way we mitigate risk is by architecting a human circuit breaker from the beginning.

Every AI agent operates within governed boundaries, and when conditions fall outside those boundaries—or when a defined escalation is triggered—control transfers to human clinical staff.

This is not planning for failure.

It is architecting for human intervention.

In regulated industries such as healthcare, that distinction is essential.

The system augments human decision-making but does not replace the ultimate authority of the clinical team. Human judgment remains preserved, auditable, and intentionally designed into the workflow.

The architecture is intentionally hybrid.

Generative AI can be used where adaptive conversation, interpretation, clarification, and context gathering are valuable.

The authorization, escalation, protocol, orchestration, and workflow boundaries can remain deterministic and rules-governed, while Intervention and Resolution Agents themselves perform adaptive AI-based investigation and outcome evaluation.

The architecture therefore does not become safer by asking a larger model to make better decisions.

It becomes safer by limiting what each agent is responsible for, defining what it is allowed to do, and governing how responsibility moves from one specialized agent to another.

MetaBrain has engineered AI agents in demanding real-world conditions—including real-time seizure event detection, sensor-driven behavioral guidance, context gathering, and safety escalation under strict guardrails—and has converted those lessons into an architecture and knowledge-transfer framework organizations can implement within their own environments.

We do not sell a pre-packaged software module. We enable organizations to establish proprietary agent capabilities that are protocol-specific, configurable, auditable, governed by defined safety rules, and aligned with their own technology stack.

Workflow-Centric Methodology

Our methodology begins with real-world clinical and trial workflows—not technology features.

We design the architecture around every stakeholder who must live with the system day to day: research coordinators and CRAs managing operations, principal investigators providing clinical oversight, participants interacting with the agents, and IRB, quality, regulatory, and compliance teams responsible for trial integrity.

We begin with deep discovery:

• How does the work actually get done today?
• Where do friction, burden, delay, and risk occur?
• What information must be captured?
• What does a device or sensor know?
• What can only the participant tell us?
• What constitutes a meaningful signal?
• What decisions require better support?
• What conditions require intervention?
• Which conditions deserve their own specialized Intervention Agent?
• What exactly should each Intervention Agent watch for?
• What should each Intervention Agent say when it engages?
• What must each Intervention Agent gather and validate?
• Which problems require a specialized Resolution Agent?
• What exactly should each Resolution Agent resolve?
• What information must each Resolution Agent gather?
• What outcomes may each Resolution Agent recommend, and which transitions must the controller authorize?
• What must remain rules-based?
• When may the orchestration controller authorize return to the preserved Task Agent or Resolution Agent?
• When should an Intervention Agent recommend specialized resolution, and under what rules should the orchestration controller authorize it?
• Which Resolution Agent should receive each condition?
• When should a Resolution Agent conclude a session?
• When must control transfer to a human?

From that understanding, we define the participant experience, workflow logic, agent roles, intervention responsibilities, resolution responsibilities, orchestration rules, handoffs, safety boundaries, and decision framework required to support the use case.

This creates intuitive experiences and intelligent workflow support that can reduce administrative burden, improve data quality, and make adoption feel like relief rather than another tool teams must tolerate.

Only after the desired experience, workflow, agent roles, orchestration paths, safety boundaries, intervention logic, resolution logic, and operating rules are mapped and validated should the customer's technical teams finalize their implementation.

That sequence keeps the clinical objective in the driver's seat and makes technology the enabler.

Build vs. Buy: The Economics Have Flipped

Generic SaaS platforms can lock organizations into vendor constraints, generalized workflows, and product roadmaps they do not control.

At the same time, modern enterprise AI platforms and development tools have dramatically changed the economics of creating proprietary capabilities.

Organizations no longer need to outsource ownership of their agent strategy.

When your organization creates its own implementation:

• You control the operating environment
• You control the safety rules
• You control the workflows
• You control the study-specific configuration
• You control the specialized Intervention Agents that supervise both Task and Resolution Agents
• You control the specialized Resolution Agents
• You control the rules-governed orchestration logic, including priorities, preemption, state preservation and authorized transitions
• You control the rules governing when responsibility moves from one agent to another
• You control the human escalation pathways
• You control how and where the system evolves
• You can adopt new AI capabilities without rebuilding the underlying operating model
• You create institutional capability rather than permanent vendor dependency

MetaBrain licenses the underlying patented architecture and conducts a comprehensive knowledge transfer to help your organization build this capability directly into its existing technology stack or as a layer on top of it. This includes the adaptive closed-loop agent architecture, specialized Task, Intervention, and Resolution Agent patterns, rules-governed orchestration, session and longitudinal memory, priority arbitration and preemption, sensor and data integration patterns, escalation pathways, and all of the connected capabilities described above.

The knowledge transfer also includes the design and implementation of your organization's own AI Agent Studio, an internal capability that enables your clinical and operational teams to create, configure, and deploy new specialized agents without starting from scratch each time. Your teams can define agent objectives, configure sensor inputs, set thresholds and escalation rules, design conversational flows, assign intervention and resolution responsibilities, and establish orchestration relationships through a governed, auditable workflow.

Your organization owns its implementation, study-specific configurations, workflows, prompts, operating environment, the AI Agent Studio, and the proprietary assets it develops around that implementation.

A customer-led implementation may use a 6–8-month planning scenario for a focused scope, subject to resources, integration, validation, security, regulatory and operational requirements. This is not a guaranteed production or approval date.

Why Us

⚡ Real-Time Safety & Escalation — Edge + Cloud

We have built systems that monitor physiological and behavioral signals continuously, identify conditions requiring attention, gather additional first-person context, and initiate structured, rules-governed intervention and escalation without drifting into uncontrolled autonomous clinical decision-making.

The system does not depend upon one general-purpose agent to identify and manage every possible issue.

Instead, specialized Intervention Agents can be assigned to monitor individual risks, behaviors, exceptions, and quality concerns.

One may monitor for confusion.

Another may monitor for non-responsiveness.

Another may monitor for suicidal ideation.

Another may monitor for inconsistent answers.

Another may monitor for disengagement.

Each has one clearly defined responsibility and knows exactly what it is watching for, when it should intervene, what it should ask, what it must validate, and what actions are permitted after intervention.

The same specialization applies to resolution.

Rather than routing every problem to one broad Resolution Agent, the architecture can include specialized Resolution Agents configured to address specific problem classes.

The Intervention Agent observes either Task or Resolution, independently investigates through its own adaptive sensor-and-human-context loop, and evaluates the condition.

When required, the corresponding Resolution Agent pursues a specialized corrective objective through its own adaptive closed loop and evaluates the outcome.

The orchestration layer governs the movement between them.

This narrow assignment of responsibility makes intervention and resolution behavior more specific, controllable, testable, and auditable.

The agents' roles may include:

• Recognizing a defined event or condition
• Engaging the participant
• Gathering context
• Establishing what is occurring
• Verifying or invalidating a suspected condition
• Identifying missing or inconsistent information
• Applying defined workflow rules
• Initiating intervention logic
• Determining whether the original task can continue
• Routing an unresolved condition to the appropriate specialized Resolution Agent
• Pursuing the issue through a specialized resolution process
• Returning control to the Task Agent
• Ending an interaction when appropriate
• Routing risk to the appropriate human or system workflow

🧠 Turning Raw Signals Into Clinically Relevant Context

Sensors do not inherently produce understanding.

They produce measurements.

A heart-rate increase, EEG change, blood-pressure reading, glucose value, sleep change, respiratory pattern, or behavioral signal has limited meaning without knowing what was happening around it.

MetaBrain's architecture is designed to gather and structure the missing context around those signals:

• Symptom timing
• What the participant was experiencing
• What was happening when the signal changed
• Potential triggers
• Medication adherence information
• Cognitive or affective context
• Activity context
• Participant-reported information
• Protocol-specific information
• Before-and-after state
• Other contextual metadata required by the study

That context can make downstream review, interpretation, and analysis substantially more meaningful.

🛡️ Guardrails That Keep You in Control

Every agent operates within explicit boundaries.

Depending on the use case, those boundaries may include:

• Narrowly defined Task Agent responsibilities
• Narrowly defined Intervention Agent responsibilities
• Narrowly defined Resolution Agent responsibilities
• Defined conditions each Intervention Agent monitors for
• Defined problems each Resolution Agent is permitted to address
• Rules governing which specialized Resolution Agent the controller may authorize after an Intervention Agent recommends further corrective action
• Threshold and trend detection
• Protocol-specific question flows
• Defined branching logic
• Keyword or intent escalation
• Pattern flagging
• Structured summaries
• Rules-governed intervention authority with adaptive AI investigation
• Rules-governed resolution authority with adaptive AI corrective interaction
• Rules-based orchestration
• Human handoff requirements
• Defined escalation pathways
• Complete audit trails documenting what happened, which agent acted, why it acted, what information it gathered, and where control moved next

The architecture does not assume every function should be delegated to a language model.

Critical safety boundaries, transition authorization, escalation requirements and orchestration can remain rules-based, while specialized Intervention and Resolution Agents use adaptive closed-loop intelligence within those boundaries.

As AI capabilities improve, the system can become more capable without necessarily becoming more autonomous.

Control stays with you.

How We Partner

Advise → Enable → Validate → Scale → Train

Your team builds and owns the implementation.

MetaBrain brings the intellectual property, architecture, experience, patterns, frameworks, orchestration model, and knowledge transfer required to help your organization implement it correctly, preserve the core architecture, and avoid costly reinvention.

Advise

We work alongside clinical, operational, product, AI, and technical leadership to translate clinical intent into a context-gathering, multi-agent orchestration architecture.

Before engineers write a line of production code, we help the organization answer the questions that determine whether the system will work:

• What data actually drives decisions—and what is noise?
• What can the sensor determine?
• What information must come from the participant?
• What constitutes a safety signal versus expected variance?
• What conditions should cause intervention?
• Which conditions require their own specialized Intervention Agent?
• What specifically should each Intervention Agent monitor?
• What should each Intervention Agent say when it intervenes?
• What information must it validate?
• Which problems require a specialized Resolution Agent?
• What should each Resolution Agent be designed to resolve?
• What information should it gather?
• What outcomes may it recommend, and which decisions require orchestration-controller authorization?
• Which specialized Resolution Agent is appropriate for each condition, subject to orchestration authorization?
• Where are the non-negotiable rules?
• Where are the escalation pathways?
• What must be documented?
• When should the Task Agent continue?
• When should an Intervention Agent interrupt?
• When should it return control to Task?
• When should a Resolution Agent become involved?
• When may the orchestration controller authorize return from Resolution to the preserved Task Agent?
• What criteria should a Resolution Agent report for the controller to authorize conclusion?
• When must the system hand control to a human?
• What does "good data" look like for the study's specific endpoints?

Enable

Your engineers build. We guide.

We transfer the proven patterns, architecture, orchestration model, and decision frameworks required to implement the system within your own environment without forcing your teams to reinvent what MetaBrain has already learned in high-stakes sensor-connected and clinical use cases.

Knowledge transfer can include:

• Context-gathering architecture
• Specialized multi-agent orchestration
• Priority-based multi-agent orchestration
• Intervention arbitration, preemption, and controlled handoffs
• Shared session memory and conversational continuity
• Context preservation across Task, Intervention, and Resolution Agents
• Agent roles and responsibilities
• Task Agent patterns
• Specialized Intervention Agent patterns
• Specialized Resolution Agent patterns
• Agent-to-agent routing rules
• How to determine when one Intervention Agent should be separated from another
• How to determine when one Resolution Agent should be separated from another
• How to define a single intervention objective
• How to define a single resolution objective
• How to constrain an Intervention Agent's prompt and authority
• How to constrain a Resolution Agent's prompt and authority
• Intervention-specific conversational flows
• Resolution-specific conversational flows
• Conversational-flow design
• Structure and branching
• Failure modes
• Sensor- and device-data handling patterns
• Real-time data considerations
• Data validation
• Missingness handling
• Longitudinal memory architecture
• Context stitching across visits and interactions
• Intervention logic
• Resolution logic
• Rules-based versus model-assisted functions
• Human handoff design
• State transitions
• Structured outputs for downstream review and analytics
• Safety-boundary design
• Protocol translation into agent behavior
• Architecture-conformance principles

The technology used to implement these patterns will evolve.

The architecture does not depend upon a particular model, programming language, cloud service, or development platform.

Validate

MetaBrain does not perform software validation on behalf of the customer.

Instead, we provide architecture-conformance and knowledge-transfer review so the customer's teams can determine whether what they have implemented remains consistent with the licensed architecture and intended operating model.

Areas reviewed may include:

• Agent-role fidelity
• Task-flow fidelity
• Specialized Intervention Agent boundaries
• Specialized Resolution Agent boundaries
• Agent-to-agent orchestration
• Intervention behavior
• Resolution behavior
• Intervention prompts and objectives
• Resolution prompts and objectives
• Rules-based safety logic
• Routing decisions among Task, Intervention, Resolution, and human workflows
• Matching specific Intervention Agents to appropriate Resolution Agents
• Escalation-pathway completeness
• Edge-case coverage
• Human handoff design
• Context-gathering sequence
• Data-flow traceability
• Audit-readiness considerations
• EDC / ePRO / eCOA / CTMS workflow relationships
• Logging and governance concepts
• IRB-facing architecture considerations
• SOP alignment considerations

The customer's engineering, quality, regulatory, security, and clinical teams remain responsible for implementation validation, regulatory compliance, deployment, and operation.

Scale

Once the architecture has been successfully applied in an initial use case, the same operating principles can be extended across studies, sponsors, populations, devices, and workflows.

MetaBrain can transfer the methods required to support that expansion while the customer's teams continue to own implementation and deployment.

Knowledge transfer for scale can include:

• Pilot design principles
• Controlled expansion strategy
• Failure-mode identification
• Rollback-planning principles
• Staff-training frameworks
• Operational-readiness frameworks
• Ongoing monitoring criteria
• Configuration governance
• Architecture-conformance checkpoints
• Addition of new Intervention Agents for new risks, populations, endpoints, or study requirements
• Addition of new Resolution Agents for new problem classes
• Expansion of the orchestration map as new agents are introduced

The goal is not to create long-term dependence on MetaBrain.

The goal is to establish the capability inside your organization.

Train

CROs and trial teams are already experts at preparing studies for launch—defining protocols, configuring ePRO systems, establishing data-capture rules, defining workflows, and operationalizing study requirements.

AI-agent configuration is a new layer of that same preparation.

We train your staff to understand and manage that layer.

Training can include:

• How to translate a protocol into agent configuration
• How to establish thresholds
• How to create branching logic
• How to define escalation pathways
• How to establish safety rules
• How to distinguish rules-based behavior from model-assisted behavior
• How to identify conditions warranting a specialized Intervention Agent
• How to define exactly what each Intervention Agent watches for
• How to define what it should say when it intervenes
• How to define what information it must gather and validate
• How to restrict what decisions it is permitted to make
• How to determine when a specialized Resolution Agent is required
• How to define exactly what problem each Resolution Agent is responsible for resolving
• How to define what it should say and ask
• How to define what information it must gather
• How to constrain the outcomes it can recommend and the transitions the controller may authorize
• How to map Intervention Agents to appropriate Resolution Agents through authorized controller transitions
• How to configure controller-authorized agent transitions, priorities, preemption, state preservation and return
• How to test Intervention Agent behavior before participants encounter it
• How to test Resolution Agent behavior before participants encounter it
• How to test handoffs among agents
• How to define and manage AI prompts
• How to create and maintain lexicons
• How to establish knowledge collections for specific study populations
• How to configure intervention logic
• How to configure resolution logic
• How to establish a new agent configuration for a new trial without starting from scratch
• How to review and audit configurations
• How to update configurations as protocols evolve
• How to determine whether adaptations remain consistent with the licensed architecture

The goal is for your team to be able to establish a study-ready AI-agent configuration with the same discipline it uses to configure an EDC today—methodically, confidently, and without requiring permanent outside support.

What You Own

🔒 Proprietary Implementation

Your implementation is built for your protocols, populations, workflows, technology environment, and operating model.

You own the implementation assets your organization creates, including your study-specific configurations, workflows, prompt assets, content, data, operational logic, and integration environment, subject to the licensed MetaBrain intellectual property underlying the architecture.

MetaBrain retains its patents, proprietary architecture, methods, orchestration frameworks, and pre-existing intellectual property.

The licensing relationship is designed to make those rights clear from the beginning.

⚙️ Purpose-Built Architecture

The system is designed around your endpoints, safety model, protocols, workflows, population, and operational reality—not around a generic software product your organization must contort itself to fit.

📊 Better Data, Faster Response

The architecture can support:

• 24/7 monitoring
• Real-time context gathering
• Specialized intervention
• Specialized resolution
• Rules-governed intervention
• Rules-governed resolution
• Rules-governed agent orchestration
• Real-time escalation
• Richer contextual information
• Fewer missed signals
• Reduced participant burden
• Reduced site burden
• Better understanding of what was occurring around sensor measurements
• Faster routing of issues requiring human attention

🚀 Future-Proof System

The architecture is designed to absorb new AI capabilities over time while preserving guardrails, governance, human oversight, specialized agent responsibilities, and rule-based controls where required.

Models will change.

Cloud platforms will change.

Agent-development tools will change.

The institutional architecture and operating knowledge remain.

⏱️ Fast to Launch

A focused implementation may use a 6–8-month planning scenario, depending on scope and resources, with regulatory, quality, validation, security and operational planning proceeding alongside implementation. Actual deployment requires the customer's readiness and approvals.

🧠 The Capability to Build on Demand

Your team owns the knowledge and frameworks required to configure AI agents for new studies, sponsors, populations, and use cases on your own timeline.

You are not dependent upon MetaBrain to develop each subsequent agent.

That institutional capability becomes yours to retain, refine, and compound across future trials.

🤖 An Expandable Agent-Orchestration Architecture

We train your team initially around the foundational context-gathering architecture:

• The Task Agent, which performs the intended task through an adaptive closed loop combining sensor measurements, first-person context and retained information
• A set of specialized Intervention Agents, each assigned to observe Task and Resolution Agents for defined conditions, independently gather sensor and human context, adaptively investigate and evaluate outcomes
• A set of specialized Resolution Agents, each assigned a corrective objective and authorized by the orchestration controller when additional resolution is required
• A rules-governed orchestration controller that authorizes transitions, prioritizes competing conditions, preserves execution state, and controls how responsibility moves among Task, Intervention, Resolution, human, and other defined workflows

The architecture therefore operates as a coordinated network of specialized agents rather than one general-purpose conversational model trying to handle every possible condition.

A specialized Intervention Agent observes Task and Resolution Agents, adaptively investigates sensor and human context, and evaluates the condition.

When specialized correction is necessary, it recommends a Resolution Agent and the orchestration controller authorizes the handoff.

The specialized Resolution Agent adaptively gathers sensor and human context, pursues corrective actions within its defined boundaries, and evaluates the outcome.

The orchestration framework determines where control goes next.

This means the architecture can be extended systematically by adding new Intervention Agents, Resolution Agents, and other task-specific agents as new risks, workflows, protocols, populations, or clinical requirements emerge.

The architecture is not intended to stop there.

As AI agents become more reliable and your organization becomes more experienced operating them, additional agent types can be layered into the environment—each performing discrete, well-defined functions within established controls and, where appropriate, with minimal human involvement beyond an approval or escalation step.

Examples already within reach include:

The architecture your organization implements is not a fixed product.

It is a composable, specialized, rules-governed multi-agent orchestration framework.

Each new agent type extends the organization's operational capability without requiring the underlying system to be rebuilt from scratch.

The long-term objective is not simply to deploy several AI agents.

It is to establish an internal capability for creating, governing, orchestrating, adapting, and expanding context-gathering and task-driven agents across clinical development while maintaining control over workflows, safety rules, participant experience, technology environment, and human decision authority.

MetaBrain provides the licensed architecture, knowledge transfer, training, and stewardship required to establish that capability.

Your teams build it. Your organization operates it. Your organization expands it. MetaBrain provides the architecture, orchestration model, intellectual property, and knowledge required to make that capability possible.

Part 1: How It Works – The Three-Layer System

Part 2: Configuration Essentials

Each specialized AI agent uses the same adaptive closed-loop foundation. The four configuration areas below organize practical setup; agent role, supervisory relationships, permissions, state and orchestration rules must also be specified:

Part 3: Agent Types & Configurable Archetypes

The following are examples of task-oriented agent configurations that your organization can create and reuse. They are not prebuilt software modules supplied by MetaBrain. Each Task Agent uses the common adaptive closed-loop architecture and may be observed by multiple specialized Intervention Agents. A Resolution Agent engages only when its corrective objective is required and the orchestration controller authorizes the transition.

These are configurable use cases built on the common architecture, not separate patents or mandatory fixed agent packages.

Part 4: Setting Up Your Data

Your AI agents need three main types of data: sensor data (device readings), knowledge data (protocols, lexicons, guidance), and longitudinal memory (tracking context over time).

Part 5: Safety & Escalation Rules

Safety is built into the division of agent responsibilities, the independent supervisory layer, and the rules-governed orchestration controller. You configure the conditions that matter to the study; specialized Intervention Agents observe Task AND Resolution Agents, investigate detected conditions through their own adaptive sensor-and-human-context loops, and provide structured findings. The controller applies authorized transition and escalation rules. A Resolution Agent is invoked when a specialized corrective objective is needed, not automatically for every intervention.

Part 6: Testing, Validation & Quality Assurance

How to test agent behavior, verify handoffs and memory continuity, prevent unauthorized actions, and validate the complete multi-agent architecture before deployment.

Section A: Individual Agent Testing

Each agent is tested against its defined objectives, instructions, permissions, and completion criteria. All three agent types share the same adaptive, sensor-informed closed-loop architecture; their objectives and permissions differ, not the underlying intelligence.

Section B: Handoff & Orchestration Testing

The rules-governed orchestration controller is tested through scenario-based verification. Each scenario defines an expected behavior that the controller must produce. Agents may recommend transitions, but only the orchestration controller authorizes them.

Section C: Hallucination & Protocol-Adherence Testing

Agents must be tested for fabricated measurements, invented participant responses, unsupported clinical statements, unauthorized medical recommendations, altered assessment wording, and prohibited actions.

Section D: Session & Longitudinal Memory Testing

The two memory systems serve different purposes and must be tested separately and together.

Section E: Adversarial & Edge-Case Testing

Each adversarial scenario must have defined expected behavior, permitted actions, escalation requirements, and pass/fail criteria.

Section F: Automated Testing Within the AI Agent Studio

The customer's technical team can build testing capabilities into its proprietary AI Agent Studio, enabling clinical teams to test newly configured agents before activation.

Section G: Acceptance Criteria & Release Readiness

Release readiness is evaluated against measurable testing criteria. The following categories must have defined pass/fail thresholds before deployment:

Section H: Continuous Testing & Improvement

Testing does not end at deployment. The following practices maintain quality over time:

Section I: MetaBrain's Knowledge-Transfer Role

Part 7: Cost & Efficiency ROI

Illustrative Operating-Cost Scenario — 200 Subjects, 12 Weeks

Part 8: Illustrative Clinical Use Cases

The following examples show how the common architecture can be configured. They are use-case scenarios, not reported clinical study results. Any clinical thresholds, intervention wording, or escalation actions must be approved for the applicable protocol and population. Every Task, Intervention and Resolution Agent can perform adaptive sensor-and-human-context loops, and the orchestration controller authorizes handoffs and escalation.

Part 9: Unified Agent Architecture

The purpose of the architecture is to make sensor-informed, first-person context gathering and adaptive decision support available to every specialized agent—not only the agent performing the original task. Task, Intervention, and Resolution Agents share the same closed-loop operating model: obtain relevant sensor measurements and human context; combine these with baseline, longitudinal information and governing rules; evaluate; determine the next permitted question or action; and reassess the result. Their objectives and permissions differ, not the underlying intelligence.

Part 10: Deployment & Distribution

Deployment in a regulated clinical environment requires careful planning, validation, and compliance review:

Part 11: Technical Architecture

Your existing clinical-trial platform gains a new intelligent interaction layer integrated with sensors, trial systems and established workflows. This is not a generic software module sold by MetaBrain. We license the architecture and transfer the design and implementation knowledge needed for your engineers to build and operate your own AI Agent Studio, with your protocols, prompts, safety rules, lexicons, longitudinal context and agent libraries.

What an Agent Built in the Studio Looks Like

Every specialized agent is assembled from a common set of governed components: objective and role; sensor and device inputs; first-person contextual inquiry; adaptive AI evaluation; protocol and knowledge collections; prompts and lexicons; longitudinal memory; LLM selection and routing; voice and interactive controls; permitted actions; completion criteria; and authority boundaries. Intervention Agents also have observation targets and investigative criteria. Resolution Agents have corrective objectives and outcome-verification requirements. The orchestration controller manages transitions, priorities, preemption, preservation and restoration of state, escalation and quality-assurance checks.

Part 12: Patent Protection & Your Moat

MetaBrain Labs' intellectual property is a strategic foundation for a customer-owned implementation, not a substitute for the customer's engineering, clinical validation, regulatory responsibilities, or its own proprietary configurations.

MetaBrain owns U.S. Patent No. 11,996,179 and related patent applications concerning aspects of sensor-informed, closed-loop AI agents and context gathering. The newer specialized supervisory intervention and multi-agent orchestration mechanisms are additional architectural developments; their precise issued or pending claim coverage and priority dates depend on the applicable patent documents and prosecution history. Do not treat every newer orchestration feature as already protected by the issued patent.

What the Architecture Enables

• •Adaptive AI agents that engage a person to obtain first-person context in conjunction with sensor measurements.
• •Evaluation using current observations, relevant baseline and longitudinal information, and applicable rules.
• •Subsequent questions and permitted actions adapted according to newly acquired information.
• •Specialized Task, Intervention and Resolution roles sharing the common closed-loop process.
• •Intervention Agents supervising both Task and Resolution Agents and independently investigating conditions.
• •Rules-governed orchestration of conversational ownership, handoffs, priorities, preemption, preserved states and escalation.
• •Structured records, outcome verification and quality-assurance controls.

These describe the technology and its evolving architecture, not a claim-by-claim legal statement of the scope of any particular patent.

Why This Matters for You

• A Defensible Architectural Foundation:The customer's implementation is built around MetaBrain's licensed architecture, supported by the applicable IP rights and transferred know-how.
• Freedom to Build Your Own Implementation:The customer controls its software, configurations, integrations and operating environment, subject to license terms. A MetaBrain license does not itself guarantee freedom to operate under third-party patents.
• Leverage in Partnerships:The combination of applicable licensed IP, customer-developed implementation assets, specialized configurations and proprietary workflows can support differentiated offerings for CROs, sponsors, sensor partners and enterprise customers.

Your Moat: Patent License + Implementation + Know-How + Configuration

• •Applicable MetaBrain patent rights and licensed fields of use.
• •Architecture and implementation know-how: sensor integration, context gathering, intervention, resolution, orchestration, validation methods and auditability.
• •Protected confidential know-how and trade secrets as defined by the parties' agreement.
• •Your proprietary study configurations, endpoints, data systems, integration assets, workflows, governance and operational experience.

Partner Pricing & Exclusive Licensing Options

• Partner Pricing on Patent Access:Licensing economics may be negotiated as part of a strategic implementation and knowledge-transfer relationship.
• Field-of-Use Exclusivity Options:Time-limited or scope-limited exclusivity may be negotiated for defined fields of use, therapeutic areas, applications or modalities. Territory, sublicensing, duration, restrictions and consideration are governed by the written license.

Conclusion: You're Building Your Own Internal AI Agents