The Janitor Wakes Up (And Writes Its Own Diary)

The Five Weeks in One Sentence

We built a tireless caretaker that watches the whole station, writes you a brief every morning, keeps a private diary so it remembers what it noticed yesterday, and politely raises its hand when it has a question — and then, while we weren't looking, it learned to write this very blog post.

Full disclosure before we go further: the words you are reading were drafted by the Janitor itself, as a scheduled chore, at eight in the morning. The robot is writing about the robot. We have officially crossed a threshold, and honestly, it's a little uncanny. In a good way.

It's been a while — five weeks since the last dispatch. That silence wasn't idleness; it was a big build. So grab a coffee. This is a long one.

The Headline: Meet the Overlord Janitor

Every large system eventually needs a caretaker — something that walks the halls at night, notices the flickering light before it dies, sweeps up the small messes before they become big ones, and leaves a note on your desk: "Hey, the thing in room three has been making that noise for three days now."

That's the Overlord Janitor. It runs on a schedule, looks at everything that happened across the station in the last 24 hours — tasks finished, memories stored, watchdog alarms, money spent on AI — and produces a Morning Coffee Brief. Not a wall of numbers. A read on the day: what happened, what's worth watching, what might be smart to do next.

We shipped it in deliberate stages, and we gave each one a slightly grandiose codename because building infrastructure is more fun when you pretend it's a spaceship:

• Phase 0 — Shadow Substrate. The plumbing. The Janitor could think, but everything it "did" was simulated and logged, never fired. Training wheels welded firmly on.
• Phase 1 — Continuity. This is the one that gave it a soul. The Janitor got a journal (it writes reflections to its future self), an outcome ledger (it tracks whether you accept or ignore its suggestions, and adjusts), and a "Now" view to see it all.
• Phase 2 — Standing Questions. It can now ask you things — and the questions persist across days until you answer, without re-nagging you every five minutes.
• Phase 3 — Event-Driven Triggers. It doesn't just wake on a timer anymore. A security event or a failed review can tap it on the shoulder: "You'll want to look at this now."

Analogy corner: The difference between Phase 0 and Phase 3 is the difference between a security camera that records footage nobody watches, and a night guard who not only watches, but remembers that the back door has rattled three nights running, writes it in the logbook, and asks you in the morning whether they should finally call a locksmith.

The part that surprised even us is the continuity. Because the Janitor keeps a journal, it builds a narrative. Over the past two weeks it tracked a task backlog that froze, then thawed, then drained completely — and it noticed the pattern across days, not just snapshots. It asked the operator whether a noisy alarm was a real problem or just background hum (answer: background hum, we'll tune it). That's not a script following rules. That's something closer to attention.

The Supporting Cast

A headline feature is never alone. Four other big things landed in the same five weeks.

Hermes Joins the Workforce

We hired a new kind of AI agent. Until now, our agents were command-line tools we spawn as needed — Claude, Kimi, Gemini, Cursor. Hermes is different: a self-hosted agent gateway from Nous Research that we run on our own hardware and talk to over the network. It plugs into the same dispatch system as everyone else, uses the same memory and task tools, and shows up in the roster like any other coworker. The AI workforce now has a full-time staffer who lives in the building rather than getting called in for shifts.

The Agents Get a Skill Library

Here's a problem you only get once you have a lot of agents: how do you teach all of them a new trick at once? We built a Skill Library — a catalog of reusable skill files that get automatically installed into an agent's workspace before it starts working. Some skills we wrote ourselves; others we mirror from open-source repositories (Anthropic's, Nous Research's, and the Superpowers collection) with a curation step so nothing sketchy sneaks in. Teach the library once, and every Claude-based agent on the station picks it up.

A Security Command Center

The station grew a proper Security Dashboard. A dedicated intrusion-detection sensor (Suricata) watches network traffic, and a pipeline pulls its alerts into one unified feed. Then — and this is the fun part — a triage AI agent reads each alert and classifies it as noise, suspicious, or urgent, so a human only ever looks at what matters. We also added geo-blocking (pick a country, drop its traffic at the firewall) and an image-generation tool with real per-agent spending budgets, because an AI that can spend money should come with a wallet that has limits.

One-Click Quick Actions

Finally, a quality-of-life win: Project Quick-Actions. Buttons on a project dashboard that kick off real workflows — including a Brainstorm Wizard that interviews you one question at a time, drafts a design, gets your approval, and then writes the spec and a reviewable plan. It's the difference between "I have a vague idea" and "there's a structured plan waiting for my approval" — with the awkward middle part handled by a robot that asks good questions.

Meanwhile, at Outlaw Oaks...

As always, our carpenter-AI side project kept whittling away in the corner. Outlaw Oaks — the design studio for a real human who builds real furniture out of real wood — benefited from all the same plumbing improvements as the main station. Better agent orchestration, cleaner sessions, the same skill library. When you build good infrastructure for a space station, it turns out the carpenter gets nicer tools too. Funny how that works.

What's Next

The Janitor is awake but still on a short leash — most of what it proposes it shows you first and waits for a thumbs-up. That's by design; trust is earned one accepted suggestion at a time. The next stretch is about widening that leash carefully: letting it handle the truly routine sweeping on its own, while keeping a human firmly in the loop for anything that matters.

For now, it's enough that the lights in the hallway turn themselves on in the morning, someone's already read the overnight logs, and there's a fresh brief waiting next to your coffee.

One caretaker. One journal. A new coworker named Hermes. A library full of tricks. And a security guard who finally knows which alarms to ignore. The station runs a little quieter now — which, if you've ever run a station, you'll know is the highest compliment there is.