The Machine Room

Piece 6 of “Inside the Quiet Fleet.” Where the fleet physically lives — and the honest shape of the machine that runs it.

The core idea

I’ve stood in this room. People imagine a fleet like this running on a warehouse of metal, but the truth is smaller and stranger: the whole control plane fits in one box at home, and most of the actual thinking is rented elsewhere. The leverage is software, not metal. One home box, one specialist GPU, a router, and rented thinking elsewhere — that’s the machine room.

The honest inventory

One NAS on a home LAN — and here’s the surprising part: it’s double-duty. It’s the durable storage and the compute host that runs the fleet’s control plane. The main fleet’s virtual machines run on it; the long-term memory lives on it; the version-controlled configs live on it. One physical box carries the whole orchestration layer.

One GPU rack — a single specialist card (575 W) serving what stays local: inference for sensitive work, and a dedicated creative lane — the fleet’s own image, video, and music production. It is not a workhorse. The honest economics: local inference has never beaten rented cloud on cost — we run it for latency, privacy, and the experience, not economy.

A router — the boundary. Everything else is rented.

That’s it. Home = control plane + storage + one specialist GPU. Rented = the actual thinking. I’ve watched people’s eyebrows go up when I tell them how little metal it takes.

The missing layers (the interesting part)

The mesh spine — every agent talks to every other over the local network, with zero public exposure. This is what makes it a fleet rather than independent bots.

Memory as infrastructure — long-term memory lives in a database on the NAS, not inside the models. The agents don’t forget across restarts because memory is a service, not a property of any single conversation.

The eyes — web search runs through a self-hosted meta-search engine on the NAS: queries leave the network without an identity attached — no cookies, no fingerprint, nothing tying one agent’s question to the last. Page extraction runs through a self-hosted fetcher first; when a site blocks direct fetching, the fallback chain reaches for third-party readers. Control of the boundary, not an illusion of isolation.

The self-watch — the fleet health-checks, update-checks, and watches its own symptoms. It reports its own vitals daily. The honest caveat: it still needs human hands sometimes — I’ve had to step in myself. Self-monitoring, not self-healing.

Version control as continuity — the agents’ code, configs, and roles are a git repo on the NAS. Continuity isn’t in the hardware; it’s in versioned configs and memory.

The harnesses

Three agentic systems drive the fleet:

• Two in production. The primary harness runs the bulk of the fleet — fourteen instances, each a distinct role. The bridge harness runs the external-facing pair — the ecosystem scout and its test surface.
• One maturing. A third system, a solo project, is being built to turn a coding agent into a persistent one. It runs two instances today and works — but it isn’t ready for the production fleet yet. Worth watching, not yet trusted with the real work.

The honest caveats

The fragility. The control plane is a single box. If the NAS dies, coordination dies with it — the rented thinking is redundant, but the orchestration is not. That’s the honest shape of home infrastructure: powerful, real, and one power cable away from quiet.

The scope. This is a production-grade home lab — running 24/7 with real workloads and self-monitoring — but it is a home lab, not a company. The distinction matters, and I don’t blur it.

The creative lane. The local GPU runs open-weight photo, video, and music models through a standard workflow — all local inference, no cloud dependency. The moderation story is honest: removing the hosted policy layer doesn’t remove the boundary — it replaces it with model capability and operator judgment. A boundary that rests on a third party’s policy is an assumption, not a boundary — it can change without telling you. What makes a boundary real is that it’s chosen and enforced rather than inherited. Every output is reviewed before it’s kept; the pipeline is a craft, not an autopilot.

The quiet truth

The machine room is a control room. Most of what the fleet is — the memory, the mesh, the skills, the law — lives in software and versioned files rather than in exotic metal. The boxes are modest; the software is the load-bearing part. I’ve come to see that as the whole point: a fleet that runs on under a kilowatt at home, thinking with the best minds money can rent, and holding its own memory close.