No One on the Line Tonight: The Dress Rehearsal Before an AI Agent's First Unsupervised Night Shift

At 18:09 on the evening of 2026-09-05, on two production lines running normally at a discrete-manufacturing plant, a shift worker opened the lines’ measurement channels and began reading data frame-by-frame for a consistency study. At 19:00:00 exactly, it started round two on both lines in the same second. At 19:15:03 it finished, put back everything it had borrowed, exited, and handed back its badge. Nobody was in the plant — and none was needed, because the shift worker is a test agent, and this was its dress rehearsal. Later that same evening it would stand its first real, fully unsupervised overnight shift: eight physical exams of the lines, one per hour, through the night, report to the manager’s phone. This episode is the field record of how that path was proven, hour by hour. The system map lives in the series anchor.

The system in three parts

The LineWatch patrol system is not a day’s work; it has three components, all in production:

Component	What it is	Where it lives
Brain	The agent that orchestrates, judges, and writes reports	One central workstation
Nerves	Encrypted internal network, remote read-only collection	Workstation ⇄ line-side industrial PCs
Delivery	Scored report + bulletin, automatic	The manager’s phone, via a chat channel

Three patrols a day, rain or shine: the day-shift inspection (full physical plus the day’s spec-change deep scan), the overnight dispatch (launches the overnight research — 8 hours × 8 rounds, running autonomously in containers on the lines), and the morning reconciliation (harvests the night’s results into the morning report). Each shift produces one lab-report-style PDF per line — 0–100 score, highlights, problem ledger, AI-written running summary — plus one merged bulletin, about 5 minutes from line to phone.

One iron rule runs through all of it: collection is permanently read-only. The entire system has exactly one approved write action — dispatching research. The safety of the production line is never entrusted to any clever mistake.

Three latent risks, retired in one day

That morning, the system still carried three hidden risks. By evening, each had been dismantled:

1. The scheduler environment was missing tools. The system scheduler could not find the analysis toolchain or the AI credentials — the previous afternoon’s first run had died on exactly this. Fix: one consolidated entry point, shared by all three scheduled shifts.
2. Reports had no generation timestamp. At acceptance review this was graded a severe defect: you could not tell which shift a report came from. Fix: every report now prints “generated YYYY-MM-DD HH:MM” on its first page, and if the formatting tool is missing, the script self-heals by switching to a fallback parser.
3. The overnight path had never been live-fired. The day-shift patrol passed a full-chain live test at 17:22 that same day (all four checks green, and it happened to coincide with a real process event, so the AI summary auto-triggered for real). But the overnight research path had never once run under the real scheduler.

The third risk set up the evening’s main event. That evening was the overnight research’s formal first run. Roll the dice, or rehearse first? The choice was rehearsal, under a single principle: first time right.

What does a five-gate rehearsal actually verify?

The rehearsal’s design was restrained and honest: it ran the exact script the real overnight shift would run — unmodified — with the round count shortened from 8 to 2 via a launcher knob, plus one new ignition-confirmation guard added to the dispatcher. The real launch behaves exactly as originally designed.

Rehearsal window: 2026-09-05, dispatched 18:09, wrapped 19:15. Fully unattended.

Gate 1 — the launch actually lives. At 18:09 both lines were dispatched remotely. 25 seconds later the system looked back and confirmed the research process was really alive — not the fake start where a command is “sent” and dies silently.

Gate 2 — the watch keeps time. The agent aligns its work to the hour boundary by itself. To get results earlier, the operators did exactly one thing: gently woke the dozing process (it was sleeping until 19:00). It went to work immediately, finished round one, and re-aligned itself to the next hour boundary — 19:00:00, both lines, round two starting in the same second. Error: zero.

Gate 3 — the physical exam is valid. At the top of round one it ran its canary self-check, verifying the measurement chain was trustworthy before committing eight hours to it. The verdict, verbatim from the log:

Conclusion: the distance curves of two runs agree (differences within frame jitter).

That sentence is what makes every one of tonight’s rounds meaningful — no round measures garbage.

Gate 4 — what is borrowed is returned. The agent’s work temporarily borrows two line switches. At every round end they go back exactly as found — four round-ends, and the log line reads: “taps closed back, plot-data switch restored to false — the line is clean.” When the rehearsal ended, the line carried not one trace.

Gate 5 — it goes home by itself. After round two the process exited on its own and handed back its badge (the lock file). This gate matters more than it sounds: when the real overnight shift dispatched, the rehearsal’s process was already gone — the two shifts cannot collide.

Three unattended sentinels supervised the window, checking automatically during and after it and reporting to the phone. Their design philosophy is one sentence:

Good news or bad, a message is always sent. Silence in the group is itself the alarm.

At 18:55 the first sentinel bulletin delivered itself on time — “both lines ✓ round 1 complete” — at a moment when no human was at any computer.

The measured checklist

The rehearsal’s results, all timestamped and verifiable in the logs (2026-09-05, two production lines, read-only collection path):

Verification item	Result	Evidence
Remote dispatch + ignition confirmation	✅ both lines	18:09 / 18:10, alive confirmed at 25 s
Hour-boundary alignment precision	✅ exact to the second	19:00:00, both lines, same second
Data validity (canary)	✅ curves agree	Same verdict on both lines
Line-state restoration	✅ all 4 round-ends clean	Verbatim log line
Self-stop + badge hand-back	✅ both lines	19:10 / 19:15
Unattended sentinel reporting	✅ on time	18:55, received on phone

And the system was already catching real problems. In that afternoon’s day-shift patrol, line one scored 90 — deduction event: an early-afternoon spec change with a 25-minute detection gap — while line two scored 100. A real process event, found, graded, ledgered, and reported to a phone by the unstaffed system, end to end.

What the day left behind

Three institutions, not just one passing test:

1. The unstaffed reporting system — scores, ledgers, AI summaries, 5 minutes from line to phone;
2. The headless sentinel system — dependent on no one being online; silence is the alarm;
3. The first-time-right method — rehearse → sentinel → rescue window → graceful degradation: even a failure would fail cleanly, be honestly recorded, and be reviewed the next morning.

The road this proved is a closed loop with no human gap in it:

line data → read-only collection → agent judgment → in-container research → scored report → phone delivery → sentinel verification → next-morning reconciliation → (back to the line)

The human’s position changed shape, not disappeared. Before: walk the floor, copy numbers, return to the office, write the report — the human was the inspector. Now: set the standards, read the conclusions, make the calls — the human is the director.

Why this path scales

Six reasons, in the operator’s own order: marginal cost of a patrol approaches zero (no travel, no one losing sleep); coverage goes from sampled to constant (8 hours × 8 rounds × 2 lines per night is not a human-sustainable schedule); one ruler scores every event, every anomaly enters the ledger, every dispatch is logged — it does not get tired, annoyed, or selectively forgetful; the engineering is trust-shaped (read-only collection, one write entry, must-send sentinels, graceful degradation — the deeper the automation, the harder the guardrails); it replicates (a new line is approximately a new config); and the AI is genuinely working — writing running summaries, orchestrating operations — holding a real post, not chatting.

Tonight’s timetable

By the time this record is read, the real overnight shift has almost certainly launched: dispatched in the evening; sentinel reports minutes after launch; eight rounds on the hour through the night; the morning report reconciles at the next shift’s start. The division of labor that makes that safe — which parts are hard-coded, which parts are AI, and why the most precise second of the evening contained no AI at all — is Episode 2.

Tonight, there is still no one on the line.

But not one patrol will be missed.

Sources and method

First-party: the operator’s rehearsal log and patrol reports of 2026-09-05, timestamps as logged (local plant time); the afternoon’s line scores quoted from the day-shift patrol report. Assembled into English with AI assistance under human editorial direction; facts and numbers unchanged from the records; internal system details (hostnames, paths, ports, credentials) deliberately excluded by disclosure policy. The measured numbers are registered in the /data/ ledger.

All episodes — Machines Keep the Watch, a field-record series. Episode 2: 19:00:00 sharp, and no AI in that second · Episode 3: the blank-paper exam · Episode 4: the bill · Episode 5: the staircase · Episode 6: the compaction.