MetaBrain Labs Inc. (Alexandrea Day)

MetaBrain Labs appreciates the opportunity to provide feedback on FDA’s discussion paper concerning generative AI-enabled medical devices.

Our attached submission describes an approach to agentic AI that we believe is particularly relevant to FDA’s consideration of safety, hallucinations, autonomy, clinical decision-making, and the integration of physiological sensor data.

MetaBrain Labs began developing its closed-loop technology using a rules-based architecture, subsequently incorporating specialized AI agents while retaining supervisory control over their responsibilities, interventions, handoffs, and decision-making boundaries.

Importantly, our agents are designed principally to gather and validate context, not independently diagnose, prescribe, or treat. Our architecture orchestrates narrowly focused task, intervention, and resolution agents under a separate supervisory controller, enabling adaptive patient interactions without granting generative AI unrestricted authority.

We believe this distinction is important to FDA’s regulatory considerations. Agentic AI does not necessarily imply autonomous clinical decision-making. Properly architected, it can improve patient engagement, integrate physiological sensor information, support more informed clinical interactions, and potentially improve outcomes while maintaining clearly defined safety boundaries.

Our attached comments provide further technical detail and recommendations for evaluating these systems based on their intended use, architectural safeguards, and system-level performance.

We welcome the opportunity to contribute our experience and engage further with FDA as it considers the evolving regulatory framework for agentic AI.

MetaBrain Labs
U.S. Patent No. 11,996,179
Pending U.S. Patent Application No. 19/366,192

MetaBrain Labs — Public Comment to the FDA

Docket: FDA-2026-N-7874
Subject: A Rules-Based, Closed-Loop Architecture for Safe Context-Gathering Agentic AI

MetaBrain Labs appreciates the FDA's examination of generative AI-enabled medical
devices, particularly the challenges of hallucinations, excessive autonomy, clinical
decision-making, contextual reliability, and the orchestration of multiple AI agents.

Our experience offers a somewhat different perspective because our technology was not
originally developed to enable autonomous AI diagnosis or treatment. It was
developed to improve the quality, relevance, and continuity of interactions through
closed-loop contextual information gathering.

We began with a rules-based architecture and subsequently incorporated increasingly
sophisticated AI capabilities while preserving the rules, supervisory controls, and defined
responsibilities governing the workflow.

Our intellectual property includes granted U.S. Patent No. 11,996,179 and pending
continuation-in-part application 19/366,192, with claimed priority extending to our
September 9, 2021 provisional application.

We believe this architecture provides an important example of how agentic AI can improve
patient engagement and potentially clinical outcomes without requiring unrestricted AI
autonomy.

1. Our agents are designed primarily to gather context, not make clinical decisions

An important distinction in evaluating agentic AI is understanding what the agents are
actually intended to do.

MetaBrain Labs' architecture is principally designed to orchestrate specialized agents that
gather, evaluate, and validate contextual information during an ongoing interaction.

These agents are not inherently intended or authorized to independently diagnose
diseases, prescribe medications, or determine treatments.

Instead, their responsibilities may include identifying missing information, asking clarifying
questions, recognizing changes in patient circumstances, incorporating relevant
physiological sensor information, and determining whether a defined condition warrants
further attention.

The architecture separates these activities into narrowly focused responsibilities:
 • Task agents conduct the primary interaction, such as a patient assessment,
questionnaire, or engagement activity.

• Intervention agents monitor narrowly defined conditions or contextual changes
that may warrant interrupting the primary task.

• Resolution agents gather and validate additional information, address the
identified issue within their authorized scope, or initiate escalation.

• A separate supervisory controller governs agent priorities, activation, interruption,
handoffs, escalation, and return to the original task.

The agents do not independently determine their own authority or take unrestricted control
of the interaction.

The objective is not to replace clinical judgment. It is to improve the information
available to patients, clinicians, and other authorized decision-makers while
maintaining continuity and control of the interaction.

2. Context gathering can include physiological sensor data

We particularly welcome FDA's attention to the importance of contextual information,
including data obtained from physiological sensors.

Our architecture was designed to incorporate contextual information from multiple
sources, potentially including patient responses, longitudinal observations, clinical
records, wearable devices, and physiological sensors.

For example, an ongoing patient interaction may be informed by changes in physiological
measurements, patient-reported symptoms, or discrepancies between what a patient
reports and what available sensor data indicates.

A specialized agent can gather additional context, seek clarification, or initiate a defined
escalation without independently rendering a diagnosis.

Importantly, sensor data does not have to initiate the interaction. Agent engagement
may be initiated by a schedule, patient input, a sensor event, or another defined trigger.
Sensor information may simply enrich the context of an interaction already underway.

We encourage FDA to recognize this distinction between using sensor information to
improve contextual understanding and using sensor information to autonomously make
consequential clinical decisions.

These activities may have materially different intended uses and risk profiles.
3. Rules-based orchestration keeps AI agents within narrow boundaries

Our architecture began with rules-based control, and that foundation remains central to its
operation.

Rather than allowing generative AI to determine its own objectives, select unrestricted
actions, or govern the overall workflow, we separate the intelligence of individual agents
from the authority to coordinate them.

Each specialized agent has a defined responsibility. The supervisory controller manages
which agent is engaged, which conditions take priority, when an interruption is appropriate,
and when the original task should resume.

For example, during a patient-reported outcome assessment:

1. A task agent conducts the assessment.

2. An intervention agent identifies a defined contextual condition requiring attention.

3. The supervisory controller interrupts the primary task and engages the appropriate
resolution agent.

4. The resolution agent gathers and validates additional information, resolves the issue
within its authorized scope, or escalates it.

5. The supervisory controller determines the appropriate next workflow step, including
returning to the original task with relevant context preserved.

This closed-loop process allows the interaction to adapt to changing circumstances
without granting the conversational AI unrestricted decision-making authority.

Our governing principle is simple: AI agents may interpret and gather information, but
they do not independently govern the workflow or expand their own authority.

4. Address hallucinations by controlling their consequences

We do not believe hallucinations can be eliminated simply by instructing generative AI
models to be accurate.

Instead, we have approached the problem architecturally.

Specialized agents are limited to defined responsibilities. Context gathering provides
opportunities to obtain missing information rather than infer it. Information validation and
resolution are distinct from the primary conversational task. Supervisory control governs
transitions and authorized actions.
When information is missing, contradictory, outdated, or uncertain, the appropriate
response may be to gather additional context, request clarification, or escalate—not to
generate an unsupported conclusion.

This separation is particularly important in healthcare.

An AI agent may produce an incorrect interpretation, but that interpretation should not
automatically become an authorized clinical action.

We believe FDA should evaluate not only the frequency of hallucinations but also whether
the architecture can prevent unsupported outputs from resulting in unsafe decisions or
actions.

5. If diagnostic or treatment capabilities are introduced, rules-based authority
remains essential

Although our context-gathering agents are not inherently designed to diagnose, prescribe,
or treat, the architecture could support future applications involving clinical decision
support or regulated medical-device functions.

In such applications, we would preserve the same foundational approach: explicitly
defined decision boundaries, rules-based supervisory control, appropriate clinical
validation, and human oversight or authorization as required by the intended use and
associated risks.

AI could enhance the patient experience by improving communication, gathering more
complete information, identifying relevant contextual changes, and supporting timely
engagement.

However, adding diagnostic or treatment-related functionality would require separate
evaluation of the clinical function, its authorized decisions, and its safety and
effectiveness.

The purpose of incorporating AI would remain to enhance patient understanding,
engagement, and outcomes—not to confer unrestricted clinical authority on a
language model.

6. FDA should evaluate agent orchestration and intended use, not simply model
accuracy

We encourage FDA to distinguish among systems that gather context, systems that provide
clinical recommendations, and systems that autonomously execute clinical actions.
Not every multi-agent healthcare interaction should be presumed to involve autonomous
diagnosis or treatment. Regulatory oversight should reflect the actual intended use,
functionality, and risk.

Where an agentic system falls within FDA's medical-device jurisdiction, we recommend
system-level testing that evaluates:

• Authority boundaries: Can any agent exceed its defined responsibilities or bypass
supervisory controls?

• Handoff integrity: Is the correct agent engaged with the necessary context
preserved?

• Context validation: Can the system distinguish verified information from
assumptions, stale data, or conflicting information?

• Hallucination containment: Can unsupported outputs produce unauthorized
consequential actions?

• Intervention and resolution: Are relevant conditions recognized, prioritized,
resolved, or escalated appropriately?

• Return to task: Can the original interaction resume accurately after an intervention
without losing important information?

• Auditability: Can the system reconstruct the sequence of agent activities,
decisions, and supervisory transitions?

Testing should include competing interventions, contradictory sensor and patient-reported
information, missing context, attempted unauthorized actions, and extended interactions
with multiple handoffs.

Acceptance criteria should be predefined and proportionate to the clinical risks of the
intended application.

Conclusion

We believe FDA's evaluation of agentic AI would benefit from recognizing an important
architectural distinction.

Agentic AI does not necessarily mean autonomous clinical decision-making.

It can also mean a carefully orchestrated system of narrowly focused agents that gather
context, recognize changing circumstances, validate information, intervene when
appropriate, resolve or escalate defined issues, and return control to the original task.
MetaBrain Labs began with rules-based closed-loop technology and evolved toward
context-gathering multi-agent AI while preserving that foundation of constrained authority
and supervisory control.

Our architecture is intended to support more informed, responsive, and personalized
patient interactions, including the integration of physiological sensor data where
appropriate.

We encourage FDA to distinguish these context-gathering applications from autonomous
diagnostic or treatment systems and to consider rules-based orchestration, constrained
agent autonomy, and closed-loop validation as important elements of safe agentic AI.

The goal is not to give AI more authority. It is to use AI's intelligence to gather better
context, improve patient engagement, and support better outcomes while keeping
authority and decision-making appropriately controlled.

MetaBrain Labs welcomes further dialogue with FDA regarding these architectural
approaches.

Alexandrea Day, CEO

MetaBrain Labs Inc.