The delivery pipeline · runs on this site

Agents build. I stay editor-in-chief.

A pipeline of agents that builds, tests and ships software on its own, with me as architect and editor-in-chief. It is not a demo, because this site was built this way. What I learn here about reliability and control goes straight into client work.

Every issue becomes its own branch and an ordinary pull request.

Four quality gates an agent cannot skip or talk its way past.

An agent never merges by itself, and never deploys.

01 · From issue to pull request

How does the pipeline work?

I describe the work, the system builds it and checks its own output, and I approve the result. That is the whole loop, in four steps.

A ruled index card on a small easel, with a red tab in the corner.

Step 01 · Linear

I write the issue.

An ordinary Linear issue written to a fixed spec template: what has to be built, what is off limits, and how I will judge it.

A small wind-up toy robot with a folder under its arm and a red key on its back.

Step 02 · The agent

An agent picks it up.

A small background process hands the issue to a Claude Code agent, which opens its own branch and builds the work.

A turnstile with three arms, one of them red.

Step 03 · Quality gates

The work gets checked.

Tests, code style, the front-end build and a review pass: checks the agent cannot skip or talk its way past.

A letter tray with a stack of paper and a single red sheet on top.

Step 04 · Pull request

I review and merge.

The result lands as an ordinary pull request that explains its choices. I stay the editor-in-chief; the agent never merges.

Who does what.

What the agents do

Read the issue and break down the work.
Write the code and the tests that go with it.
Run the quality gates and fix their own failures.
Deliver a pull request that explains its choices.

What an agent may not do

No merging and no deploying: I review the pull request, I merge it, and production is never a direct output of the pipeline.
No work outside the issue: what is not in there does not get built.
No skipping, disabling or softening a quality gate, and no deleting tests to turn the build green.
No adding or replacing dependencies without my approval.

02 · Four quality gates

How does the work get checked?

Every gate runs automatically, on every branch. If one of them is red, the work never reaches my desk. The boundaries live in the issue and in the repository, not in a good conversation afterwards.

Gate 01

Test suite

A clipboard with a checklist, the top box ticked in red.

The full test suite runs on every branch. Red means no pull request.

Gate 02

Code style

A small typewriter with a sheet of paper in it and red ribbon spools.

Laravel Pint guards the formatting, so a diff is about substance and not about whitespace.

Gate 03

Front-end build

The assets are built the way production builds them. If the build breaks, the work stops there.

Gate 04

Review pass

Reading glasses with red frames on a folded sheet of paper.

A separate agent reads the work against the issue before I ever see it.

Want to see this run on your codebase?

A half-hour call costs nothing. You will hear what it takes to make this work on your codebase, and what does not.

Book a call

03 · Deliveries from the pipeline

What can you check?

These are real merged pull requests from the pipeline. The repository is private, so there are no links; the numbers, dates and line counts come straight from the git log. Every line passed the same four gates.

artisan-at-work/artisanatwork 5 merged

EN-site staat op v2 vanuit copy-en-v2.md ART-237

#81 · merged on 27 Aug 2026 · gates green

+657 −149

Navigatie, contact-chips en redirects staan op de v2-sitemap ART-236

#80 · merged on 27 Aug 2026 · gates green

+603 −397

Lab toont Cranely en maakt de eigen agent-productiestraat zichtbaar ART-235

#79 · merged on 27 Aug 2026 · gates green

+1166 −113

Over en Werk staan op v2: ik-vorm, loopbaan vanaf 1997 ART-234

#77 · merged on 27 Aug 2026 · gates green

+1126 −176

Platforms & CMS vervangt de drie losse dienstpagina's ART-233

#76 · merged on 27 Aug 2026 · gates green

+564 −1064

Private repository; numbers, dates and line counts come straight from the git log.

04 · From proof to offer

And on your side?

This pipeline is my strongest proof that agents can deliver reliable work, provided the boundaries live in the system. Two ways to do something with it yourself.

A short conveyor belt with a checkpoint gate and a red lamp on top.

The Build

Want this for your own team?

Agent workflows are a building block of The Build: bounded work a system carries out on its own, with checkpoints where a human approves.

See The Build

An open toolbox with neatly arranged tools and a red screwdriver on top.

Agent Delivery Kit

Set it up yourself?

The Agent Delivery Kit is this same pipeline as a download: working files plus playbook, set up in one afternoon.

View the kit

05 · What people want to know

Any questions?

The questions that always come, with short answers. If yours is not among them, a half-hour call is the fastest route.

Does an agent ever merge by itself?

No. Every delivery is an ordinary pull request that I read and merge. Production is never a direct output of the pipeline.

What happens when a gate is red?

Then the agent fixes its own failure and runs the gate again. If that does not work, the work never reaches me; it stays on the branch until it is right.

Which agents and models do you use?

Claude Code as the agent, with a small background process that hands over issues from Linear. The gates are simply the test suite, Pint, the build and a separate review agent.

Can this run on our codebase too?

Usually yes, if there is a test suite that says something and a repository the boundaries can live in. If that is missing, it starts there, and that is a Build.

A conveyor belt with three boxes, the red one just rolling off.

Who uses this

Uses it every single day.

Everyone talks about agents building software. I want to show what it takes to make that reliable: not a smarter model, but gates that cannot be bypassed and a person who keeps the final word. This site is the proof that it works, and the place where I find out where the friction still is.

Daniel Plomp · AI Architect & Builder

Want to see this run on your codebase?

Describe in two sentences what your team builds and where the work gets stuck. You will hear what it takes to let agents build reliably alongside your team.