Operating notes · Note 03
May 18, 2026 · 7 min read
Agentic practice
Agentic

What zero full-time
developers actually requires

Zero is an operating description of Steller after the pivot. It is not a claim that the software runs unattended, and it is not a payroll figure.

↓ What the number depends on
01
The metric

Zero is a headcount
of one kind of role.

The home page and Built From the Rubble use the same figure: zero full-time developers to operate Steller. I am a non-technical founder. There is no engineering team and no CTO in that account. The tools named as the remaining engineering capacity, once the outsourced team was offline, are Cursor, Qwen, and Claude. The operator the essay names for day-to-day work after handover is Qwen Coder.

That sentence does not say the company has no people, no vendor, and no human decisions. It says the platform is not staffed with full-time developers. Unpublished headcount, contractors outside that sentence, and payroll are not in the record, so this essay does not estimate them. The Complete Story’s account of the rebuild period is personal, not a compensation table: no salary, no team, no guarantee. I will not turn “no salary” into a dollar cost of founder time.

The figure is also dated to “today” in the way the essays use it: after the 27-day pivot, with the system handed to Qwen, and with Bede — the digital wallet of Zain Sudan — as the published first partner integration. It is not a description of 2024, when Steller was founded with an outsourced team.

0
Full-time developers in the published operating claim
27d
Pivot from microservices to the agent-ready monolith
$672
Cursor spend for that pivot
~$15
Expected daily AI spend after handover
02
Architecture

The number is false
on the old shape.

Zero full-time developers was not available on the system the outsourced team left. That stack, as the rubble essay records it, was ASP.NET Core 9, PostgreSQL, RabbitMQ, and Vue.js, with the Bamboo Card Portal API as the vendor. The data model was live on April 2, 2025, at 36 tables. Agents could fix the service in front of them and break two others they could not see. My line in that essay: every time I asked the agent to fix an order flow bug, it changed the right service but broke two others it couldn’t see.

The replacement that makes a solo operator plausible is the one the essay actually shipped. Fat controllers hold the golden flow in one path: wallet debit, Bamboo vendor call, PIN return, webhook dispatch. Background work is a named Hangfire job, PlaceBambooOrderJob, including the wallet rollback when Bamboo fails. The two Vue.js 2.6 dashboards became React 18, TypeScript, and Vite. The API response envelope is mandatory. Permission tables were flattened. The rubble essay’s deploy is docker compose up -d. The backend AGENTS.md keeps that command and refuses to treat its success as application readiness.

The Complete Story states the dependency in one sentence. The AI did not rescue the platform. The architectural redesign made the platform rescuable by AI. Zero full-time developers requires that redesign. It does not require a smarter prompt on top of invisible RabbitMQ consumers.

The rebuild shipped with zero critical defects in production. The method behind that claim is not published. I will not invent a test plan to make the sentence sound more rigorous than it is. The claim stands as published, and the method stays unknown here.

03
Docs and locks

Someone still has to
forbid the clever move.

Architecture without a first file puts the orientation cost back. The published drop is from fifteen to twenty minutes to about two: read AGENTS.md, open the controller. Qwen reads that file, runs tests before done, and logs changes. The cache hit rate published next to that practice is 95.2%. Docs written once are what later sessions read.

The locks are the other requirement, and they are Steller’s real AGENTS.md practice, not a metaphor. CORS allows only explicit origins, never a wildcard in production. Hangfire WorkerCount must be 2 for Bamboo and must not be increased. PlaceBambooOrderJob does not retry; it refunds on failure and swallows a refund throw. Bamboo calls always send an Idempotency-Key. JWT_KEY comes from the environment and is never hardcoded. Catalog validation does not coalesce nulls. RequireBusinessClaimsMiddleware is never bypassed.

The release lock is the human part of “zero developers.” Never push directly to main, and never run git push origin main. The root file’s chain is feature branch, tests pass, pull request, merge, auto staging deploy. Production deploy and rollback are manual GitHub Actions operations with human approval. Feature branches are feat/... or fix/..., and the backend is pushed before the root submodule ref. Closed-loop CI, when the ask is to run until green, is gh workflow run, gh run list, and gh run view --log-failed — not a person pasting logs. After a config change, ./scripts/smoke-staging.sh runs only once health is ready. The file’s example probe is http://127.0.0.1:6091/api/health. Compose coming up is not that readiness. The onboarding smoke is signup, then GET /api/wallet/me with a non-zero trial balance, then POST /api/orders returning 202, then a poll to Completed. dotnet build Steller.sln stays at zero errors. Tests are fixed only in test files. Order.cs, Brand.cs, ApplicationDbContext.cs, and migration files are not deleted or recreated to clear an error.

Three files named in the hardening week — QWEN_PROTOCOL.md, HANDOVER_TO_QWEN.md, CHANGELOG_AI.md — are not available to quote. Zero full-time developers cannot be defined as “follow those documents” until their bodies exist in the record. The operable substitute is the AGENTS.md pair and the rule that changes get logged.

04
Who remains

The roles the essays
actually name.

What remains, stated only from published roles:

01
The founder
A non-technical founder in Mississauga. The human who approves production, and the person the essays hold responsible for the redesign.
02
Qwen Coder
The named operator after handover. Reads AGENTS.md, runs tests before done, logs changes. Not a full-time developer on a payroll the essays never publish.
03
The partner
Bede, the digital wallet of Zain Sudan, is the published first integration. A partner is not an engineering department.

Cursor, Qwen, and Claude are the tools named when the team went offline. I will not add a cast of unnamed agents, a support rota, or a fractional CTO. Advisory work I describe on the home page — partnerships, business development, growth — is a separate offer. It is not evidence of staff inside Steller.

The earlier employers in the published career are Zain Sudan, Emirates Driving Company, and Rittal Technology, then Steller. None of those lines is a hidden engineering bench for this platform.

05
The limit

The number stops
where the essay says it stops.

Fat controllers, the rubble essay says, break when a team grows past a handful of engineers, when one path must scale independently of another, or when a domain holds many flows. Steller, in that essay, has one golden flow, not forty. This essay does not claim the platform has outgrown the pattern, and it does not claim zero full-time developers will survive those exit conditions. The principle published next to the limits is narrower: design so an agent can hold the relevant flow in one context window. For a larger system the essay itself suggests domain-bounded services with explicit, synchronous interfaces — not event-driven, and not invisible.

Microservices are not called universally wrong. They are called wrong for a team of one human and several AI agents. Hiring the handful of engineers the limits chapter describes would change the constraint. I have not published that hire, so I will not describe it.

“The lesson isn’t that microservices are wrong. It’s that microservices are wrong for a team of one human and several AI agents.”
Built From the Rubble · March 2026

Zero full-time developers, as Steller publishes it, requires a visible golden flow, named jobs, one deploy command, an AGENTS.md an agent will actually read, locks the agent must not relax, a human on production, and a daily AI cost the essays put at about $15, with Qwen at about $5–15 a day once the docs exist. It does not require a myth that nobody is accountable when the smoke fails.

Next in the series
Note 04 · Operating notes
When Fat Controllers Stop Working
The exit conditions the rubble essay already stated.
Read the essay →