Marcus Holloway is a product manager evaluating BlueMirror for a partnership his employer is considering. He has spent two days with the architecture documents and one afternoon with Aiyana Whitehorse from the academic medical center who has agreed to walk him through the integration questions. The question he opens with is the one product managers always open with when an agentic system serves multiple user goals at once. What happens, he asks, when the agents disagree.
The example Aiyana picks for him is a Tuesday afternoon in Margaret’s life. The earning concierge wants to schedule a 4 p.m. tutoring session because the student in Brisbane has a window before her exam. The cognitive concierge says today is a low-capacity day because Margaret slept poorly and the morning conversation showed mild word-finding difficulty. The social connection concierge sees that 4 p.m. is Margaret’s only opportunity to call Ruth before Ruth leaves on a two-week trip. The health concierge wants Margaret to rest because her blood pressure has been trending high and the cardiologist has flagged stress as a contributing factor.
Four agents. Four legitimate recommendations. One person. One time slot. Someone has to lose.
The orchestration layer’s conflict resolution architecture determines who wins, why, and how Margaret is informed of the decision. This is not an edge case. It is the normal operating state of a system that serves the whole person rather than one domain. A system without conflicts is a system ignoring domains.
Why conflicts are normal#
A single-domain system never has conflicts. A health monitoring app does not produce internal disagreement, because all of its outputs come from the same logic operating on the same goals. A scheduling app does not produce internal disagreement for the same reason. The conflicts only appear when a single system is responsible for multiple domains that have legitimately different goals.
BlueMirror handles thirteen domains. Health recommendations conflict with earning goals when the recommendation is rest and the earning opportunity is now. Social needs conflict with rest requirements when the social opportunity is also bounded. Financial caution conflicts with quality-of-life spending when the financial advice is to wait and the quality-of-life choice is to act. The conflicts are not bugs. They are evidence that the system is actually serving the whole person rather than picking one domain and pretending the others do not exist.
The architectural choice is to embrace the conflicts and resolve them well. The alternative, which is to suppress the conflicts by reducing the number of domains, would produce a simpler system that serves the person less.
The priority hierarchy#
Not all concierge agents are equal in conflict resolution. The architecture defines five tiers.
Tier 1 is safety. Health safety, cognitive safety, physical safety. The concierge agents in this tier always win. If the health concierge says rest because the blood pressure trajectory is dangerous, rest wins. If the cognitive concierge says the person should not be making a financial decision today because cognitive capacity is impaired, the financial decision waits. The override is automatic and not adjustable by the person. Margaret cannot tell the system to ignore a Tier 1 concern, because the consent for Tier 1 enforcement was given at setup and the system treats the prior consent as authoritative when present capacity is compromised.
Tier 2 is health maintenance. Non-urgent health recommendations. The tier wins against everything except Tier 1 and Margaret’s explicit Tier 3 override. A blood pressure reading that warrants a medication conversation is Tier 2. The conversation is scheduled. The earning opportunity that would compete with the conversation is moved.
Tier 3 is the person’s explicit preference. If Margaret has said she wants to teach today despite the system’s lower-priority recommendations, the explicit preference overrides Tier 4 and Tier 5 but does not override Tier 1 or Tier 2. The autonomy of the person is honored within the bounds set by safety and health.
Tier 4 is optimization. Earning, social, purpose, financial optimization. These compete among themselves based on the person’s demonstrated priorities. The system has learned over time which Tier 4 activities matter most to Margaret. Teaching matters a great deal. Social calls matter a great deal but with rotation. Financial optimization matters less because Margaret has stable finances and does not seek to maximize them.
Tier 5 is maintenance. Home maintenance, routine scheduling. Yields to everything above.
The priority hierarchy is not hidden from the person. Margaret can inspect it. She can see why the system made the choice it did. She can override Tiers 2 through 5 with explicit instruction. The only tier she cannot override is Tier 1, and the architecture is transparent about why.
Resolution strategies#
Not every conflict requires a winner and a loser. The orchestration layer prefers resolution strategies that preserve as much of the original intent as possible.
Time-shift moves a lower-priority activity to a different time. The tutoring session moves to tomorrow. The social call to Ruth moves to this evening. The system checks each agent’s flexibility, identifies the activities with shift tolerance, and selects the shifts that resolve the conflict with the smallest disruption. Time-shift is the first resolution attempted, because it usually preserves all four goals at the cost of timing only.
Scope reduction shortens an activity to fit alongside others. The tutoring session runs thirty minutes instead of sixty. Margaret rests for an hour first, then teaches the shorter session. The reduction is calibrated to the activity’s value at different durations. The student in Brisbane benefits from a thirty-minute session more than zero minutes. Margaret’s rest at one hour benefits her more than zero rest. The scope reduction trades each activity’s full value for partial values that sum higher than any single complete activity.
Delegation shifts the logistics of resolution to the concierge agents themselves. The earning concierge notifies the student about the rescheduled session, including the rationale framed appropriately. The social connection concierge proposes the evening call to Ruth. The home environment concierge adjusts the afternoon settings for the rest period. Margaret does not handle any of this. She experiences a coordinated afternoon rather than a series of individual conflicts she has to resolve herself.
Escalation surfaces the conflict to the person when the system cannot confidently resolve it. The escalation is itself calibrated. If two Tier 4 activities are close in priority and the resolution strategies produce comparable outcomes, the system presents the choice to Margaret with a recommendation. “You have three things competing for this afternoon. Here is what I would suggest and why. Would you like to choose differently?” The escalation is rare because most conflicts resolve cleanly within the priority hierarchy. When it happens, it is not a failure. It is the system honoring the autonomy that the architecture is built to protect.
The cross-concierge event bus#
The conflicts above describe what happens when two or more concierge agents arrive at recommendations that compete. The architecture also tries to prevent the conflicts from arriving at the user-facing layer in the first place.
The mechanism is the cross-concierge event bus. When one concierge agent schedules an action, the event propagates to all other concierge agents through the H-layer. The cognitive concierge publishes “low capacity day” as an event at 9 a.m. when the morning conversation provides the signal. Every concierge agent that was planning afternoon activities receives the signal and adjusts before producing a recommendation. The earning concierge knows by 10 a.m. that today’s tutoring opportunity should be approached with a shorter session or a deferral suggestion. The social connection concierge knows that today’s calls should be brief or rescheduled. The home environment concierge knows that the afternoon ambient setting should support rest rather than alertness.
The conflict is resolved before Margaret sees any recommendation. When the system works correctly, she does not know a conflict existed.
The event bus is not a fancy abstraction. It is a structured publish-subscribe channel inside the H-layer that every concierge agent reads on every relevant decision. The events are typed: cognitive state shifts, health alerts, schedule changes, family communications, external commitments. Each concierge agent has a subscription model declaring which event types affect its recommendations. The model is built into the agent’s specification, not configured at deployment.
The latency cost of the event propagation is bounded by the H-layer’s strong-consistency commitment for preference and state changes that affect user-facing behavior. The cost is paid because the alternative, eventual consistency for these events, would produce visible contradictions between concierge agents that arrived at recommendations from inconsistent state.
When the person overrides#
Margaret says “I want to teach today” despite the low-capacity assessment. The system does not argue. It adjusts.
The session is shortened from sixty minutes to forty-five. A check-in is scheduled at the thirty-minute mark to assess whether the session should continue. The student in Brisbane is informed that the session may end early. The cognitive concierge increases its monitoring during the session, ready to surface a graceful exit if the cognitive signals deteriorate.
The override is honored because the architecture is built to honor autonomy. Margaret is not a child to be protected from her own choices. She is a person whose system serves her. The system serves her by adjusting the conditions under which her choice can succeed, not by blocking the choice itself.
The exception is Tier 1. If the cognitive assessment shows confusion severe enough to be a safety risk during a teaching session involving financial information, the system explains why it cannot proceed and offers alternatives. The boundary is narrow and well-defined. It is not “the system thinks teaching is unwise today.” It is “the system has determined that proceeding with this specific activity creates a safety risk that exceeds the threshold for autonomous override.” Margaret is told what the threshold is and why it has been crossed. She is offered alternatives that respect both her preference and the safety constraint.
The narrow exception is what allows the broad respect for autonomy to remain credible. A system that overrides the person frequently, even with good intentions, becomes a system the person works around. A system that overrides only when safety requires it, transparently and rarely, becomes a system the person trusts.
Cross-references#
The Company of One (BMT-01.SYN). The synthesis where all thirteen agents work together. The conflict resolution described here is what makes the synthesis function in practice.
The Human Agency Scale (BMT-04.01). The autonomy framework that governs overrides. The Tier 1 exception is anchored in the agency framework defined there.
Cognitive Capacity and Consent (BMT-04.05). What happens when cognitive capacity affects the person’s ability to override. The Tier 1 exception in this article is rooted in the consent architecture detailed there.
How a Request Becomes an Action (BMT-02.04). The single-request trace. This article extends from single-request reasoning to multi-agent conflict reasoning.
Technical Appendix BMT-02.06-A is available to partners and investors at partners.bluemirror.tech.