The Machine Grew a Brain

A self-healing control room for an autonomous fleet

The Fleet

Five dev machines ship code around the clock — someone has to keep the lights on

A foundry runs five autonomous dev machines, each building software continuously. They're registered by name in fleet-config.yaml.

The output isn't a demo. It's real scale —

word4
botslack
openM365
safeguard
universal-app

Proof of Scale

word4 alone shipped 278 features across 106 autonomous sessions in 5.5 days

Stated in the repo README and corroborated by word4's own docs. This is output no human can shadow session-by-session.

Which turns oversight itself into the real problem —

278

features shipped

106

autonomous sessions

5.5

days

~89K

lines of code

3,714

tests

445

commits

The Gap

Running them blind burns compute unnoticed — and SSH-per-machine doesn't scale to five

A silent failure keeps running until someone notices. Checking each machine by hand means five separate SSH sessions.

So the fleet grew a place to see everything at once —

A silent failure keeps burning compute until a human happens to look
Checking status means SSH into each machine, one at a time
That doesn't scale across a fleet of five

Mechanism · Visibility

The fleet grew a ~9,944-LOC control room on `127.0.0.1:8111`

An amplifierd plugin with two modes. It reads v2 fleet status (status.yaml per machine) and prefers it over v1 Docker polling — one place to see all five.

Seeing failures is step one. Fixing them is step two —

9,944

LOC (6,540 monitor + 3,404 dashboard)

8111

port on 127.0.0.1

2

operational modes

v2

preferred over v1 sources

Mechanism · Self-Healing

A self-healing agent reasons over error output instead of static regex

The env-remediation agent tries a hints catalog first, then reasons about the stack. It's capped at 3 attempts and never modifies application source.

Now the fleet fixes itself. Next, it reaches out —

Reasons over regex: reads error output rather than matching static patterns
Catalog-first: tries the hints catalog, then reasons about the stack
3-attempt cap: global limit on fix attempts
Never touches app source: only environment, dependencies, config

The Turn

Then it phones home: morning-digest-v2 emails a plain-text fleet summary

Oversight flips from human-pull to machine-push. The fleet reports on itself via a hardcoded SMTP block instead of waiting to be asked.

Which means the human no longer has to go looking —

1

Summarize the fleet

morning-digest-v2.py builds a plain-text fleet summary email

2

Send via SMTP

smtp.gmail.com, port 587 — from amplifierschillace@gmail.com

3

Land in the inbox

Delivered to ramparte@gmail.com — "Digest sent"

Payoff

The human watches from a phone — and the fleet minds itself

The 400px desktop chat drawer collapses to a mobile bottom sheet below 768px. Self-monitoring, self-healing, and phone-home mean supervision fits in a pocket.

No SSHing into five machines. Just a glance —

400px

right drawer on desktop

<768px

collapses to a mobile bottom sheet

1

phone instead of five SSH sessions

The Pattern

Give an autonomous fleet a brain — and human attention becomes the scarce resource you actually spend

Monitoring, self-healing, and phone-home let a fleet manage its own attention budget. Autonomy scales only when the machines mind themselves.

Sources

Sources & Research Methodology

Source repo: ramparte/amplifier-bundle-dev-machine Active

Data as of: HEAD 10fd25c (2026-06-18); 164 commits; first commit 889a909 (2026-02-25)

Research performed (git / gh / grep):

git remote -v && git log -1 --format='%H %ci' && git rev-list --count HEAD — repo, HEAD, 164 commits
find dashboard/src -name '*.py' | xargs wc -l — 15 files / 3,404 LOC dashboard
find monitor -name '*.py' | xargs wc -l | sort -n — 6,540 LOC monitor
grep -n -E 'word4|botslack|openM365|safeguard|universal-app' monitor/fleet-config.yaml — five machines
sed -n '5p' README.md — word4: 278 features, 106 sessions, 5.5 days, ~89K LOC, 3,714 tests, 445 commits
grep -n 'amplifierd.plugins|8111|Two operational' dashboard/README.md — plugin, two modes, port 8111
grep -n 'reasoning over regex|3-attempt|not modify' agents/env-remediation-agent.md — self-healing contract
head -12 monitor/morning-digest-v2.py; sed -n '36,43p' monitor/fleet-config.yaml — digest + SMTP block
grep -n '@media' dashboard/src/dev_machine_monitor/static/index.html — 400px drawer / <768px bottom sheet

Gaps: The "2 AM silent failure burns compute" framing is narrative motivation, not a quoted repo string. Contributor breakdown was not enumerated via git shortlog; only config-embedded emails/paths were observed, so no per-author commit share is asserted.