Series
The Orchestration Layer
The thirteen concierge agents are what the person experiences. Beneath them, thirty-one infrastructure agents and thirty small language models do the actual work. Series 02 is the engineering explanation of how the two connect: how one slow reasoning layer directs many fast execution layers, how context is routed without wasting it, and how a multi-agent system is kept from contradicting itself.
BMT-02.01
The Brain and the Hands
The core architectural decision: one reasoning brain directing many specialized hands. The H-layer holds coherence; the L-layer delivers speed. This article explains why both …
BMT-02.02
The Thirty-One
The thirty-one infrastructure agents that power the thirteen concierge agents, organized by domain, with autonomy defaults and deployment preferences that reflect the risk and …
BMT-02.03
The Thirty Models
Thirty small language models are the target portfolio, but none run at launch. This article describes the portfolio, the five constraints that forced decomposition, and the …
BMT-02.04
How a Request Becomes an Action
A single request traced from spoken language to delivered response. Eleven steps, five models, 333 milliseconds. The article walks every step, names every latency, and specifies …
BMT-02.05
What the System Learns
P-RLHF learns Margaret's preferences separately from Dorothy's and maintains individual models for each. This article describes the learning cycle, the cold start, cross-domain …
BMT-02.06
When Agents Disagree
When four concierge agents arrive with competing recommendations for the same time slot, the orchestration layer resolves the conflict through a five-tier priority hierarchy and …
Synthesis