David Okonkwo is the systems architect on a Phoenix-area home health agency that has agreed to deploy BlueMirror into eighty households over the next year. He is reading the agent inventory because he needs to know what runs where, who calls what, and where his agency’s existing care management software fits. He opens the document expecting either a marketing list (thirty-one named features dressed up as agents) or an undifferentiated technical specification (here are thirty-one identical objects with different parameter values). What he finds is neither. The thirty-one infrastructure agents are organized by domain, each with a defined autonomy default, a deployment preference, and a clear boundary on what it does and does not do.
This is the workforce. The thirteen concierge agents in Series 01 are the public face. The thirty-one infrastructure agents are who the concierge agents direct. They organize by domain because domain expertise cannot be cleanly separated from domain context. A generic task executor cannot manage medications, because medication management is not a generic task. The Medication Manager understands pharmacy workflows, drug interaction patterns, refill cadence, adherence signals, patient assistance programs, and copay arithmetic. None of that is portable to other domains. A scheduling agent does not need it. A nutrition agent does not need it. The decomposition follows the structure of the work.
Each infrastructure agent carries two attributes that David needs to know before he integrates. The autonomy default is the level of independent action the agent can take before requiring human approval. It runs from 0.25 to 0.75 on the Human Agency Scale, with the value reflecting the risk profile of the domain. The deployment preference is where the agent runs. Zone 1 only (Local Pane, in the subscriber’s home), Zone 1 primary with Zone 2 fallback, Zone 2 (Community Pane, regional node), or Zone 2 with external system integration. The preference reflects the latency requirement and the privacy constraint of the data the agent handles. Privacy-critical agents that invoke Zone 1 models (cognitive assessment, emotional state, voice processing, safety screening) run in Zone 1. Agents that require cross-domain reasoning or the full MoC context run in Zone 2. The three-zone compute architecture is described in BMT-06.03.
These two attributes are not configurable by the agency. They are set by the architecture. The agency’s deployment touches the integration surface, not the infrastructure tier.
Why thirty-one#
The decomposition follows from the concierge architecture. Each concierge needs infrastructure agents that own specific capabilities. The number is not a target. It is a count.
The health concierge needs eight because healthcare has eight distinct operational domains: medications, appointments, care transitions, symptoms, cognition, nutrition, exercise, and vital signs. Each domain has its own data model, its own external integration requirements, and its own pattern of escalation. Collapsing them into fewer agents would force each agent to reason across domains it does not need to reason across. Splitting them into more agents would create coordination overhead between agents that should be one agent.
Family coordination needs five. Memory care needs six because the population requires specialized capabilities not shared with general interaction. External integration needs five for the major external system categories the architecture touches at infrastructure depth. Blue Pane membrane needs five for the membrane functions described in Series 03. The total is thirty-one. The number is driven by domain logic. It is not driven by engineering convenience or by any preference for round numbers.
The decomposition has implications for how the agency’s existing software integrates. David’s care management system tracks appointments, medications, and care plans. It will integrate with three of the eight healthcare agents (Appointment Coordinator, Medication Manager, Care Transition Manager) plus one of the five family coordination agents (Communication Manager). The other agents he does not touch. The existing functionality of the agency’s software is preserved. The BlueMirror agents add capability around it.
Healthcare agents#
The largest domain group. Eight agents.
The Medication Manager handles reminders, refills, adherence tracking, and interaction checks. Autonomy 0.75. Deployment preference: Zone 1 primary, with Zone 2 reach for prescription database queries. The most frequently invoked healthcare agent. Multiple concierge agents call it. The health concierge calls it for the obvious reasons. The buying agent calls it for procurement timing. The nutrition concierge calls it for dietary interactions with specific medications. The high autonomy reflects the observational character of most medication management; the agent reminds, tracks, and notes, but does not change the prescription. Changes go through human approval.
The Appointment Coordinator schedules, arranges transportation, sends reminders, and confirms. Autonomy 0.5. Deployment preference: Zone 2, depending on the integration target. Lower autonomy than the Medication Manager because scheduling involves commitments that affect other people. A cancellation creates downstream consequences for the provider and for the family. The agent does not initiate cancellations autonomously; it queues them for the person’s approval.
The Care Transition Manager handles discharge planning, home services coordination, and follow-up. Autonomy 0.25. Deployment preference: cloud, because the integration surface for care transitions is large and changes frequently. The lowest autonomy in the healthcare domain because care transitions involve irreversible decisions and multi-party coordination. A misstep here costs days of recovery and possibly a readmission. The agent gathers information, builds the transition plan, and presents it for approval. Execution proceeds only after approval.
The Symptom Monitor performs pattern detection across self-reported symptoms and vital signs, generates alerts when patterns warrant attention. Autonomy 0.5. Deployment preference: Zone 1 primary. Continuous background operation. Feeds the health concierge with trend data. The agent does not diagnose. It identifies patterns and surfaces them. The diagnosis remains the physician’s responsibility.
The Cognitive State Assessor tracks orientation, detects cognitive fluctuation, monitors lucidity. Autonomy 0.75. Deployment preference: Zone 1 where available. Shared between the health concierge and the cognitive concierge. For subscribers with a Local Pane, Zone 1 is the deployment target because cognitive assessment requires sub-second response and because cognitive state data is the most sensitive data the system holds. For those subscribers, the data does not leave the home. For subscribers without a Local Pane, the assessor runs in Zone 2 (where regional coverage exists) or Zone 3 (otherwise) under the healthcare data processing agreement. The deployment substrate differs across subscriber paths; the assessment logic does not.
The Nutrition Tracker logs meals, tracks dietary adherence, surfaces deviations from prescribed restrictions. Autonomy 0.5. Deployment preference: Zone 2. Feeds the nutrition concierge and the buying agent. Lower autonomy than the Medication Manager because dietary recommendations are often contested by the person and require collaborative adjustment rather than autonomous adjustment.
The Exercise Monitor tracks activity, assesses mobility, integrates with wearable devices. Autonomy 0.5. Deployment preference: Zone 1 primary, with Zone 2 reach for cross-domain correlation. Fall detection integrates here when wearable hardware supports it. The agent does not prescribe exercise. It monitors what the person does and surfaces patterns to the health concierge.
The Vital Signs Analyst trends blood pressure, glucose, weight, and oxygen saturation, producing the time-series data the Symptom Monitor reasons over. Autonomy 0.75. Deployment preference: Zone 1 primary. Continuous monitoring. High autonomy because trending is observational, not interventional. The agent identifies a trend; the health concierge decides whether the trend warrants action; action requires the person’s approval.
Eight agents covering eight operational domains. Each with a defined boundary. Each with a defined autonomy. Each calling specific small language models from the thirty-model portfolio described in BMT-02.03.
Family coordination agents#
Five agents. Calendar Coordinator, Decision Facilitator, Communication Manager, Visit Scheduler, Care Circle Notifier.
The work of these agents is information flow management. The person served does not become the switchboard for her own care. The Communication Manager routes the right information to the right family member at the right moment. The daughter receives the weekly health summary because she has subscribed to it. The son does not receive it because he has not. The granddaughter receives a different summary, focused on social engagement, because that is what she has asked for.
The Decision Facilitator manages family decisions that require input from multiple people. The choice of a specialist, the timing of a move, the response to a hospital readmission. The agent gathers input, visualizes options, tracks consensus, and presents the result. It does not make the decision. It removes the coordination cost of making the decision well.
The Visit Scheduler manages rotations among caregivers and visiting family members. The Care Circle Notifier sends targeted alerts when the situation changes, with the targeting calibrated to each member’s preferences and information needs.
The privacy boundaries within the family are enforced here. The daughter sees the weekly health summary. The son does not. The son sees the home maintenance status. The daughter does not. The agents do not leak across these boundaries. Each family member’s view is constructed for her, and what she sees is what the person served has authorized her to see.
Autonomy across these five agents averages 0.5. Deployment is Zone 2 with external system integration, depending on which integrations the family uses. Calendar integration with Google or Outlook routes through Zone 2 to the external provider. Direct in-system family messaging runs at Zone 2 with notification delivery to Zone 1 devices.
Memory care agents#
Six agents. Orientation Assistant, Reminiscence Facilitator, Routine Anchor, Wandering Prevention, Sundowning Support, Communication Adapter.
All six deploy at high autonomy, 0.75. Five of six deploy in Zone 1 only. The reasons are tightly coupled.
The latency argument first. Cognitive support that takes three seconds is cognitive support that embarrasses. A person experiencing disorientation does not have three seconds to spare. She is already losing the thread. If the system pauses to query a regional node or an external service, the moment is lost and the support becomes a confirmation of the deficit rather than a correction of it. The Orientation Assistant must respond in well under a second. The Confusion Detector that triggers it must run continuously in the background. Both run in Zone 1. There is no other place they can run and still serve the function.
The privacy argument second. Cognitive state data is the most sensitive data the system holds. A leaked medication list is a privacy violation. A leaked cognitive assessment is a profile of mental decline that affects insurance eligibility, employment, family relationships, and legal capacity. The system’s privacy commitment is to give every subscriber the strongest privacy posture her deployment path can support. For subscribers with a Local Pane, the architecture enforces this by deploying the cognitive assessment agents in Zone 1. The data exists only on the Local Pane in the person’s home. There is no Zone 2 copy. There is no central archive of cognitive trajectories that could be subpoenaed because the data was never centralized in the first place. For subscribers without a Local Pane, the strongest available posture comes through the healthcare DPA governing Zone 2 or Zone 3 inference: no retention beyond the request lifecycle for transient queries, no use for the provider’s own model training, audit rights, breach notification. Contractual protection is weaker than architectural data residency, and the architecture distinguishes which subscribers receive which kind.
The high autonomy reflects the urgency of the support. A person showing signs of sundowning at 4 p.m. cannot wait for human approval to receive a calming intervention. The Sundowning Support agent acts. The action is logged. The family is notified. But the support is not delayed.
The Communication Adapter, the one Zone 2-eligible agent in this group, adjusts language complexity in real time based on cognitive state signals. It runs at Zone 2 when the regional node is reachable, falls back to Zone 1 when it is not. The agent’s output is reviewed by the Safety Filter (Zone 1) before delivery, which keeps the cross-zone transition transparent to the person.
External integration agents#
Five agents. Pharmacy Liaison, Provider Communicator, Insurance Navigator, Transportation Coordinator, Emergency Responder.
These are the agents that touch systems outside BlueMirror. Each requires a specific trust level from the external party, established through the Blue Pane membrane described in Series 03. The external system does not see the person’s full context. It sees only what its trust level entitles it to see.
The Pharmacy Liaison handles refill orders, copay calculations, patient assistance program enrollment, and prescription history sync. Autonomy 0.5, deployment Zone 2 with external integration. Trust level requirements: the pharmacy must be enrolled, must support the refill API, and must accept the trust attestation the membrane provides.
The Provider Communicator routes structured communications to the person’s physician panel. Appointment summaries, between-visit questions, vital sign deltas the physician asked to be flagged. Autonomy 0.25, deployment Zone 2 with external integration. The lower autonomy reflects the formality of provider communication. The agent drafts; the person approves; the message is sent.
The Insurance Navigator handles claims, prior authorizations, appeals, and benefits questions. Autonomy 0.25, deployment Zone 2 with external integration. The same logic applies. Insurance communication is consequential and adversarial. The agent prepares; the person approves; the action proceeds.
The Transportation Coordinator integrates with rideshare, paratransit, and family driver scheduling. Autonomy 0.5, deployment Zone 2 with external integration. Higher autonomy because most rides do not have downstream consequences worth the friction of human approval. A ride to the grocery store does not need a confirmation step.
The Emergency Responder is the exception that proves every rule. Autonomy 0.75, deployment Zone 1 primary with mandatory external escalation. Trust level requirement: maximum. This is the only infrastructure agent permitted to break privacy boundaries in a life-threatening situation. When a fall is detected and the person does not respond to the verification prompt, the Emergency Responder shares the person’s location, medical conditions, and emergency contact list with the responding service. The action is not negotiable. It is the architectural answer to the question: what does the system do when the person cannot consent because the person is unconscious. The answer is to treat the prior consent, given at setup, as authorization for the emergency action.
Blue Pane membrane agents#
Five agents. Context Gate Controller, Trust Scorer, Negotiation Sandbox Manager, Manipulation Detector, Audit Trail Logger.
These are not user-facing. They are the infrastructure that protects the user from the external agent world that BlueMirror operates inside. The full treatment is in Series 03. They are introduced here because the inventory is incomplete without them.
The Context Gate Controller manages what external agents can see, by trust tier and by domain. The Trust Scorer evaluates and assigns trust tiers in real time as external agents interact with the membrane. The Negotiation Sandbox Manager creates isolated environments for agent-to-agent interactions where commitments are not binding until the membrane releases them. The Manipulation Detector identifies urgency attacks, preference probing attempts, and inference extraction patterns that indicate an external agent is trying to obtain information it has not been granted. The Audit Trail Logger records every interaction with cryptographic signatures that allow after-the-fact reconstruction of what an external agent did.
These five agents deploy across Zone 1 and Zone 2 depending on their function. The Context Gate Controller, Trust Scorer, and Manipulation Detector run in Zone 1 to keep latency low for external interactions and to enforce membrane decisions at the home boundary. The Negotiation Sandbox Manager runs at Zone 2 because it requires substantial compute for multi-turn agent-to-agent negotiation. The Audit Trail Logger runs at both, with Zone 1 logging for low-latency capture of decisions made at the home boundary and Zone 2 aggregation for long-term retention and cross-domain analysis.
Cross-references#
The Thirteen (BMT-01.01). The concierge agents these infrastructure agents serve. Each concierge agent’s article in Series 01 references the specific infrastructure agents it orchestrates.
The Thirty Models (BMT-02.03). The small language models that power the infrastructure agents. This article describes the agent abstraction; that article describes the models that execute under it.
The Membrane (BMT-03.01). The full treatment of the Blue Pane membrane agents introduced in this inventory. The protective infrastructure that makes the external integration agents safe to deploy.
The Human Agency Scale (BMT-04.01). The autonomy framework that produces the 0.25 to 0.75 values cited throughout this inventory. The article defines what each level means and how the values are calibrated per domain.
Technical Appendix BMT-02.02-A is available to partners and investors at partners.bluemirror.tech.