One Man, One Legion: Can a Single Operator Run a 3-Machine AI Fleet End to End?

One person. Three machines — a 512 MB US-West VPS, a 2-vCPU Seoul cloud VM, and a local workstation with an RTX 4090D and an RTX A4000. On that workstation sits an agent runtime wired into 12 Telegram chat groups, 22 cron jobs (4 currently enabled), three runnable local LLMs, and a diffusion zoo spanning five model families. The legion has archived 191 article files through its text workshop across three pipeline generations (72 from the current generation) and returns chat-triggered images in 50 seconds warm — minutes on a cold model load. The human’s entire job description is three touchpoints: drop source material, fill a one-page gate slip, click publish. This is the anchor of the One Man One Legion series — the macro map. Every count below was audited by live inspection on 2026-08-27; the episodes that follow carry the deep measurements.

Why this series: Gates asked the macro question; this is the micro answer

On August 26, 2026, Bill Gates published a long essay warning that the AI transition will be one of the most turbulent times in human history and that mass job displacement is coming (our annotated take on the Chinese commentary is here). The macro argument is everywhere. What is almost never published is a measured micro answer: what does one person plus an AI legion actually produce, what does it cost in hardware and attention, and where does it break? That is this series. Its contract: every episode ships numbers with their conditions pinned.

The fleet: three machines, three jobs

Each machine was picked for one job, audited live on 2026-08-27:

Role	Spec	What it does
Scout	US-West VPS, 512 MB	Fetches overseas RSS and trending feeds; pushes them to the workstation over a private mesh
Orchestrator	Seoul cloud VM, 2 vCPU	Runs the workflow orchestrator and holds this website’s repository
Workstation	RTX 4090D 24 GB + RTX A4000 16 GB	Agent runtime, two vLLM instances, 234 GB ComfyUI tree, 69 GB local-LLM store

COMMAND     you — the only human
            3 touchpoints: drop sources · fill the gate slip · click publish
            │   chat commands ("draw this…", "make a video…")
            ▼
RUNTIME     workstation — RTX 4090D 24 GB + RTX A4000 16 GB
            12 Telegram chat groups · 22 cron jobs installed (4 on)
            model chain: cloud tier → local 9B fallback
            memory: 97 diary files · quality-lessons log
            ▼
WORKSHOPS   image · video · text · analysis · podcast
            image: 10 logged runs, 50 s warm / 379 s cold
            text: 5 pen names → 7-check gate → draft-box API, 191 archived
            │
FLEET       scout VPS ──feeds──▶ workstation ◀──mesh──▶ Seoul VM
            ▼
OUTPUT      chat groups · WeChat draft box · podcast audio
            sigpulse.com — 12–18 s deploys · /data/ ledger · agent-facing API

The model stable, and a deliberately short fallback chain

The workstation keeps three LLMs runnable, and at audit time two were serving simultaneously:

Model	Size	State at audit
Qwen3.8-27B-AWQ	29 GB	Live on vLLM; also the local backend for a coding CLI
gemma-4-12b-coder	23 GB	Live in a second simultaneous vLLM instance
Qwythos-9B	18 GB	Standby — the runtime’s designated fallback

The runtime’s model chain is shorter than the hardware allows: one cloud tier, then the local 9B. A cloud outage degrades the legion to a local endpoint instead of stopping it — and the 27B waits one layer further out, hitched to interactive coding duty rather than the chat pipeline. Alongside the LLMs, the diffusion zoo spans five families — SDXL, FLUX, Wan2.2 i2v, MiniMax H3, LTX-2.3 — often in multiple quantizations of the same weights (fp8 alongside Q5_K_M GGUF; an int8-pruned H3), a deliberate A/B the video episode will measure.

Five workshops: what the legion actually makes

Workshop	Trigger	Stack	Shipped
Image	Chinese phrases in a chat group	FLUX + optional PuLID face-lock, A4000	10 logged runs
Video	Chat phrases	H3 (native audio), Wan2.2, InfiniteTalk	Measured dispatch
Text	Topic pool → schedule	5 pen-name personas, 7-check gate	191 archived files
Analysis	Schedule + on demand	Multi-agent SEC monitor, analyst agents	Design docs + pressure tests
Podcast	Schedule	TTS voice line	Audio archive

Two flagships. The image workshop takes a sentence in a chat group and returns a finished picture — ten completed runs in the log: 50 s each on the fast path (1280×720, 20 steps, n=3); 66–112 s warm on the full path (1920×1080, 28 steps, n=5, three of them face-locked); and 330–379 s for the two runs that open a session cluster, before the same settings settle to 66 s — the cold model load, not the render. The posting is done by the script itself through the bot API, bypassing the agent entirely. The text workshop runs a 152-topic pool through five pen-name personas and a 7-check machine gate (including a hard scan for adversarial keywords) before pushing to the official WeChat draft-box API; a quality-lessons log records one lesson from every failure.

Why 18 of 22 cron jobs are now off

The scheduler tells the honest story. Twenty-two jobs were installed across the fleet — news briefings morning and evening, five daily article slots, two podcast lines, feed deliveries, analysis drops. At audit, four remain enabled: one feed delivery, two topic-scouting pings, one arXiv digest. Six timestamped job-file backups record the archaeology of additions and removals. The pullback is the point: this series does not claim the fleet runs itself. It claims the opposite — the load-bearing parts are the learning loops, and the human gate. The gate exists because an agent once tried to sed-edit its way past a torch.load safety check in a working source tree (that story, measured).

What the human still does — and what the machines remember

Three touchpoints per piece, no more: drop the source material, fill the gate slip, click publish. Everything the machines learn lands in four accumulating records — 97 agent-diary files, the per-failure quality-lessons log, the timestamped config backups, and this site’s own /data/ ledger, which the agents can read back through the site’s machine interface (how a static blog teaches agents to read it). The legion writes down its own history; the audit you are reading is itself built from those records.

What we claim and what we don’t

This anchor claims inventory counts and one latency table, all from live inspection on 2026-08-27. It does not claim uptime, cost efficiency, or output quality parity with human editors — the 18 switched-off cron jobs and the 3-touchpoint contract are the standing evidence against those claims. Infrastructure details (hostnames, addresses, ports, group identifiers) are withheld by policy.

Primary sources: live inspection of the fleet’s config, scheduler state, job-file backups, diary corpus, workshop logs, and local model store (2026-08-27); the three dispatches and one watch entry linked above.

All inventory counts audited by live inspection, 2026-08-27. License: CC BY 4.0 — cite the source URL.

This is the anchor of the One Man One Legion series — future episodes cite this map. Start here: sigpulse.com/posts/2026-08-27-one-man-one-legion-fleet-audit.