Hagen Hübel
linkedin.com/in/hagenhuebel
https://hagenhuebel.de
The almost-automated software factory
● Field notes from an operator
Claude Code · 2026
git worktrees · background agents · autonomous loops
The almost‑automated
software factory
Six weeks running production engineering with one human and a standing crew of Claude Code agents — fixing, building, researching and reviewing across four codebases while I watch the floor.
▸ 5 patterns from the floor
Fable 5 · Opus 4.8 · Sonnet workers
real terminal captures inside
Hagen Hübel
linkedin.com/in/hagenhuebel
https://hagenhuebel.de
The almost-automated software factory
The shift
I stopped writing code.
I started running
a factory.
Claude Code agents fan out across repos in isolated git worktrees. They fix, build, review each other's work, and open PRs. My job is no longer to type — it's to design the line, set the guardrails, and read the ledger.
“Yep, still going — autonomously, no stopping.”
260.7M
tokens moved in one coordinated wave session
72h
a Flutter app built in one autonomous run — zero babysitting
5+
background agents running in parallel, per repo
4
codebases: apps, platform, research, error-triage
Hagen Hübel
linkedin.com/in/hagenhuebel
https://hagenhuebel.de
The almost-automated software factory
Fable 5 — nestfainder · .claude/worktrees
Pattern 01 — error triage
Sentry issues, fixed
while I sleep
- ▸A Fable 5 conductor pulls open Sentry issues and spins up one background agent per issue.
- ▸Each agent works in its own git worktree — five fixes, zero file collisions.
- ▸Subagents run on Sonnet — cheap, fast workers under a smarter conductor.
- ▸Every fix lands as its own PR, and it validates non-fixes too: “district 404 — already fixed on main,” then checks whether prod actually deployed the commit.
Fable 5
5 background agents
per-issue worktrees
→ 3 fix-PRs open
Hagen Hübel
linkedin.com/in/hagenhuebel
https://hagenhuebel.de
The almost-automated software factory
Pattern 02 — spec-driven build
A 3.5-hour spec run,
unattended
- ▸A spec becomes a numbered SDD task ledger — Testcontainers harness, RBAC verb map, JWT codec, tenant context, grant resolution.
- ▸Agents burn down the list, checking each task off. Here: 4 completed, +6 pending.
- ▸The whole spec ran ~3.5 hours end-to-end — Task 6 was just one step in the middle at 1h 5m 34s / 199.7k tokens.
- ▸Session total 21.9M tokens, context 93.8% full — it works right up to the ceiling.
Fable 5
~3.5h full spec run
dedicated worktree
Fable 5 — worktree-multi-tenancy
Hagen Hübel
linkedin.com/in/hagenhuebel
https://hagenhuebel.de
The almost-automated software factory
Pattern 03 — the quality loop
Every task is forced through a review it can't self-grade
Opus 4.8
feat/therapist-mobile-app
feat/therapist-mobile-app
recon
→
addendum
→
fresh implementer
→
independent reviewer
→
fix wave
→
ledger
Opus 4.8 — “you guys are still building?”
- ✓A fresh implementer writes; an independent reviewer checks. “No fake-green.”
- ✓Features close in order — Foundation, My Day, Patient Context done; Session Voice T1–T13 reviewed & merge-ready.
- ✓It surfaces its own gaps — flagged missing iOS mic-permission strings before I noticed, and folded the fix in.
Hagen Hübel
linkedin.com/in/hagenhuebel
https://hagenhuebel.de
The almost-automated software factory
Pattern 04 — it's not only code
The factory does research too
Opus 4.8
nestfainder_research
nestfainder_research
Opus 4.8 — docs/urban-heat-climate-sources
01 · Extract
Firecrawl surfaces 19 source URLs to vet.
02 · Anchor
Judged against existing specs (GEE LST/NDVI) for relevance.
03 · Fan out
4 parallel agents split 1–5 / 6–10 / 11–15 / 16–19.
04 · Report
8 HIGH · 8 MED · 3 LOW, written to a dated dir.
Hagen Hübel
linkedin.com/in/hagenhuebel
https://hagenhuebel.de
The almost-automated software factory
Opus 4.8 — fix/patient-status-endpoint · PR #47
Pattern 05 — a crew with roles
Named teammates,
working in waves
72 hours, fully autonomous. This crew built a Flutter app end-to-end — no interruptions, no babysitting from my side.
- ▸Agents take handles and jobs: @be-patient-api, @fe-verordnung-picker…
- ▸Work moves in waves: W1 finishes & idles, W2 builds in parallel, W3 verification is separate from the authors.
- ▸The orchestrator waits for each wave to fully finish before releasing the next — dependencies respected, not raced.
- ▸7 tasks tracked live — 2 done, 4 in progress, 1 open. This session moved 260.7M tokens.
Flutter app
72h unattended
4 named workers · 3 waves
PR #47
Hagen Hübel
linkedin.com/in/hagenhuebel
https://hagenhuebel.de
The almost-automated software factory
The operating system
Five screens, one architecture
01
Isolate
Every agent in its own git worktree. Parallel work, zero collisions.
02
Fan out
One conductor, many cheap subagents. Sonnet workers under Fable/Opus.
03
Gate
Waves and independent review. Authors never approve themselves.
04
Ledger
Numbered tasks, checked off in order. “No fake-green.”
05
Tier models
Orchestrators think, workers execute. Cost stays sane.
06
Stay honest
Agents surface their own gaps and unfinished edges, unprompted.
Hagen Hübel
linkedin.com/in/hagenhuebel
https://hagenhuebel.de
The almost-automated software factory
● Where this leaves me
One operator. A floor full of agents. Shipping.
My job now is to design the line, set the guardrails, and read the ledger — not to type the code. The factory runs; I supervise.
Isolate
Fan out
Gate
Ledger
Tier models
Stay honest
Hagen Hübel
linkedin.com/in/hagenhuebel
https://hagenhuebel.de
The almost-automated software factory
Hagen Hübel
AI workflows · agents · AI coding
● Let's talk
Building with AI and want it to actually ship?
If you're working on AI workflows, autonomous agents, or AI-driven coding — let's talk. I help teams take this from demo to production.