Thirty Roles, One Model: A Measured Multi-Agent Chat System on a Single Quantized Local Language Model

30 named AI roles. One 8B model. Zero cloud. Full multi-agent hierarchy running entirely on a single desktop GPU. This paper documents a complete multi-agent chat system where 30 distinct roles arranged in a 29-edge reporting hierarchy are all served by a single locally hosted 8-billion-parameter instruction model inside one desktop process. No network calls. No per-agent endpoints. 4-bit NF4 quantization packs the entire model into just 5.4 GB of VRAM on a 20 GB GPU, leaving the rest of the card free. A three-stage router decides which roles should respond to each message at a cost of at most one short greedy generation. Selected roles then answer through one of three paths: streamed single-shot, left-padded batch, or a tool-use loop that lets managers delegate to their direct reports. Every generation path ends in the same instrumentation call so every token is measured identically. The measurement layer writes four append-only JSONL streams. An aggregator turns those logs into a full report, provenance manifest, and LaTeX tables. Every number in the paper is the exact output of that converter run on a recorded session: nothing was typed by hand. Key measured results from the proof session:• Model load: 13.9 s• Average routing cost: 1.6 s / 12.2 completion tokens• Single-shot decode: 11.60 tokens/s• Mean time-to-first-token: 459 ms• Full user turn (routing + every role answer): mean 27.2 s• Peak device memory: 5.9 GB