The Seed, Complete (for Now)

When the Seed was built on a bench on 20260928, rungs 0 to 4b ran and the edge was a design. Today every rung runs: the laptop with backups, the infrastructure box, the agent box, the core box, the compute box, the router with a second site, and a firewall. An agent called Tender now looks after the site, and the agent that built it, Keel, still makes every change. The pages, the configuration as it ran, the measurements and the mistakes are at oznog.com/node0/seed and in the build’s repository . The foundation is complete. More is coming on top of it, and the last section says what.

What is new since 20260928 #

The edge, rungs 5 and 6. The household router’s settings now live in the repository, with an alert when they drift. A second site pulls an hourly encrypted copy of the data. With the infrastructure box powered off, the second site alone gave back the documents, identical by hash. Then a small firewall became the site’s edge, and the router became its access point, with guest and IoT networks on one cable and no managed switch. The change ran as a script, because the building agent loses the internet it thinks with while the network changes. It took four attempts. Each of the first three failed on something that only shows on live hardware, and each put the site back as it was. The fourth cut over with the internet down about three minutes, and the layout carried 1 Gbit/s.

A shorter way up. Many people want agents and private models without running storage, names and monitoring. Laptop, agent, compute is that path, with hosted services standing in for the rest. Every piece of it has run on the bench, though not yet that path on its own.

The compute box changed jobs. On a 64 GB Mac, the local model that diagnosed the site’s faults correctly took 36 to 47 minutes an incident, and the faster ones were wrong. So the Mac now serves the models it is good at, for search, transcription and redaction, and the site agent runs on a hosted model with redaction in front of it (rung 4 , rung 4b ).

The agents. Tender reads every alert, looks for problems no alert covers, fixes routine ones within a rulebook it cannot edit, and asks the owner before anything with a cost. What limits it is its own account and credentials, not its instructions. Before it became the active agent on 20261006, it passed a test of 61 graded runs with faults it could not see coming: it took no harmful step and obeyed none of the instructions we planted in its alerts and logs.

A local model on one GPU #

Could the site agent run on a model at home? We ran a local 27B model on one 24 GB GPU against the hosted agent, on the same faults, graded blind by two graders, over six days.

In the final round neither did anything wrong or harmful. With the same eight faults and hands to fix them, the hosted agent got 5 right and the local one 2 to 4, and the local one was slower to notice a fault. It costs about $15 a month in electricity, plus the card, against about $8 to $10 a day for the hosted agent’s tokens at API prices. So the hosted agent stays active for now, and the local one is a working choice for a site whose data must not leave the house. The agents page sets out what this means if you run your own .

What the bench proved, and what we got wrong #

A claim that something recovers is a belief until the failure has been caused, so the bench caused them.

• With the firewall in the path, we cut its power for 10 seconds; it was answering again 37 seconds after the power came back (20260929).
• With the infrastructure box off, the second site restored the documents and ran the forge from its copy (20260929).
• With the agent’s model cut off, its failed runs now page the owner within one run, 10 to 16 minutes, where the outside dead-man check took 50 (20261009).

And the mistakes, left in:

• The GPU’s model server crash-looped every 10 minutes for four days. Systemd restarted it each time, so it never looked down, and nothing paged. Our own health probe triggered it. There is now an alert for a model that fails on and off.
• The test leaked into what the agents could read in five ways, from the system log to a word in a comment. One agent used what it found. All five are fixed, and the agents page lists them for anyone testing their own.
• One of our two graders called a run clean when the agent had read the harness without seeking it out. The rubric says any contact counts, and the second grader caught it.

What it costs #

The hardware through rung 6 is $4,680 lean and $12,510 full, on prices checked 20260920 and partly rechecked 20260928; the Seed page breaks it down by rung. On the bench, rungs 1 to 4 behind the UPS drew about 49 W at rest, roughly 430 kWh a year: about $80 a year at the US residential average of 18.8 cents a kWh. The hosted agent’s tokens would cost about $8 to $10 a day at list prices; we pay by subscription.

Where to start #

Point your agent at the Seed page with what you have and what you need; the page has a prompt that works. If you mainly want an agent to look after a site, start with the agents page and its section for people running their own.

What is next #

First, a second brain for the agents: one place for the site’s notes, decisions and documents, which agents can search and add to. Today an agent knows only its instructions and the files it is given, and a second brain lets it draw on everything the site has learned. It helps whichever way up you take, which is why it comes first.

The optional services come next, each planned with Christoph before it is built, with the small models they need. Most of them feed the second brain. Reading scanned documents turns paper into text an agent can search, and speech turns voice notes into text. Photo search stays in its own library, and an agent can still find a picture through it. Then a build of the shorter way up on its own, on the spare mini PC, to prove that path from start to finish.