Margaret’s daughter Elena lives 90 minutes away and coordinates her mother’s care from a distance, which means she spends a meaningful portion of her week on the phone confirming details that Margaret’s system already knows and that Elena’s calendar already holds. When Elena’s own personal AI was set up, one of the first things she wanted to configure was a direct connection to her mother’s system, so that the two agents could coordinate without requiring either of them to be the relay.
Two personal agents coordinating care across households is different from every other integration described in this series. The pharmacy agent, the hospital scheduler, the insurance platform: those are institutional agents interacting with a personal agent. The relationship is asymmetric. One side is the service provider; the other is the person being served. When Margaret’s agent meets Elena’s agent, neither side is the service provider. Both sides have context. Both sides have preferences. Both sides have privacy requirements. And the interaction affects a human relationship.
Why peer interactions are harder
Institutional agent interactions are asymmetric by definition. The pharmacy agent needs something from the person’s system, or the person’s system needs something from the pharmacy. One side holds context; the other holds service capacity. The trust tier system across this series has been one-directional: the person decides whether to trust the pharmacy agent. The pharmacy does not need to evaluate whether to trust Margaret’s agent. It trusts it because it is Margaret’s agent.
Peer interactions are symmetric. Margaret’s agent needs to trust that the agent claiming to be Elena’s is actually Elena’s and is accurately representing Elena’s situation. Elena’s agent needs the same assurance about Margaret’s. Both sides have context that could be shared and context that should not be. Neither side is automatically the authoritative party. The trust model must operate bidirectionally.
There is also a different kind of stakes. When a hospital scheduling agent makes an error, the corrective path runs through the hospital’s systems and institutional accountability structures. When Elena’s agent makes an error, the corrective path runs through a human relationship. A failed agent-to-agent interaction between Margaret and her daughter affects something that cannot be filed as a bug report.
Family agent architecture
Family care coordination is the most common form of peer agent interaction, and the family coordination concierge manages the person’s side.
Margaret’s family coordination concierge holds her preferences about what she shares with each family member. She wants Elena to see health summaries but not cognitive assessment scores. She wants Elena to be notified about medical appointments but not about financial transactions. She wants Elena’s agent to be able to schedule appointments on her behalf but not to modify her medication list.
Elena also has a personal AI, and Elena has her own preferences about what her agent shares with Margaret’s. Elena does not want Margaret to know she has been researching respite care options, because she does not want Margaret to feel like a burden before they have had that conversation directly. Elena wants her agent to share the family visit calendar but not her work schedule.
Two agents, two privacy configurations, two legitimate sets of boundaries within a loving relationship. Margaret’s family coordination concierge enforces Margaret’s configuration. Elena’s AI enforces Elena’s. Neither side has unilateral access to the other’s full context. The coordination happens within the intersection of what both sides permit.
When Elena’s agent asks about Margaret’s upcoming appointments, Margaret’s family coordination concierge responds based on Margaret’s configuration for Elena: yes, there is a cardiology appointment Thursday morning, at the main campus, accessibility accommodation confirmed. It does not share that the appointment is a follow-up to a recent hospitalization Margaret has not yet told Elena about, because Margaret’s configuration does not permit that disclosure without Margaret’s direct involvement.
Trust reciprocity
In institutional interactions, trust is one-directional: the person decides whether to trust the pharmacy agent. In peer interactions, trust is bidirectional. Margaret’s agent must trust that the daughter’s agent is actually Elena’s and is not an impersonation. Elena’s agent must trust that Margaret’s agent is sharing accurate information. The trust tiers apply symmetrically, but the typical starting point is higher: family agents often begin at TIER_4D or TIER_5E based on the person’s explicit configuration during setup.
The starting tier is not automatic. Margaret must actively configure it. The system does not infer family trust from a declared relationship. Trust at TIER_5E is the person’s explicit choice.
Margaret’s membrane must also verify that the agent presenting as Elena’s is actually Elena’s. The peer agent authentication protocol requires mutual credential verification: Elena’s agent presents credentials that prove it is associated with Elena’s BlueMirror instance, and Margaret’s system verifies that credential chain before any context is shared. Impersonation of a family member’s agent, to gain the access TIER_5E provides, is an attack vector the authentication protocol is designed to close.
Once authentication is confirmed and trust tiers are established on both sides, the interaction proceeds within a shared sandbox configured for peer interaction, with reciprocal exploration bounds. Margaret’s family coordination concierge can share what Margaret’s configuration permits. Elena’s agent can share what Elena’s configuration permits. The sandbox records the exchange.
Beyond family: community agent interactions
Neighbor agents coordinating shared yard care start at TIER_2B, not TIER_4D. The neighbors have a relationship, but not a care relationship, and the context shared between their agents is minimal: shared service preferences, scheduling availability, cost splitting preferences. No health data. No financial detail beyond the shared service cost.
A church community agent coordinating meal delivery for homebound members operates at TIER_2B to TIER_3C depending on the relationship history. The coordination context is limited to what is necessary: that Margaret is a current participant, that she has dietary restrictions of a specified type, that delivery is preferred on specific days. The medical reason for the dietary restrictions does not transfer to the community organization.
Community agent interactions are less sensitive than family interactions in health terms, but they introduce a trust bootstrapping problem that family interactions do not have. Family trust starts high because Margaret assigns it explicitly during setup. Community organization trust must be established through the standard evidence package process or through organizational attestation: the community organization provides a credential chain that the Trust Scorer can verify, which elevates the starting tier from TIER_1A to TIER_2B and accelerates the path to TIER_3C.
The relational stakes
When Margaret’s buying agent negotiates with a vendor agent and the negotiation fails, the consequence is that a purchase does not complete. Margaret finds another vendor. The relationship is transactional, and a failed transaction is a minor inconvenience.
When Margaret’s family coordination concierge and Elena’s agent fail to coordinate correctly, the consequence can be a missed appointment, a care gap, or Elena arriving for a visit when Margaret expected the following week. These outcomes damage a human relationship. The system is not responsible for the relationship, but it is responsible for not making the relationship harder.
This is why the peer sandbox includes escalation to both humans, not just one. In an institutional interaction, escalation goes to the person whose agent is negotiating. In a peer interaction, either agent can escalate to its own person. If Margaret’s agent and Elena’s agent reach an impasse on a care coordination decision, both Margaret and Elena are notified, with the current sandbox state visible to each. The decision comes back to the humans, which is where a disagreement between family members belongs.
Elena’s configuration resolved the connection two weeks after she set up her personal AI. The initial coordination worked without either Margaret or Elena intervening. The following Monday, Margaret’s cardiology appointment showed up in Elena’s calendar with the notes her agent had permission to see. Elena called her mother that evening not to confirm the appointment, but to talk.
Cross-References#
The Family Coordination Concierge (BMT-01.14). The concierge agent that manages family peer interactions from the person’s side.
Trust Tiers and What They Unlock (BMT-03.02). Trust tier system applied to peer agents with bidirectional evaluation.
The Social Connection Concierge (BMT-01.09). Community-level peer interactions and the social infrastructure beneath them.
Three Pools of Expertise (BMT-08.01). The Personal Circle as the peer interaction layer of the Expert Exchange Layer.
Technical Appendix BMT-03.07-A is available to partners and investors at partners.bluemirror.tech.