Edouard Tardif
FREN

A factory where agents build software, and a human stays in charge.

Since 19 July 2026, AI agents have been turning written specifications into real applications: they split the work, write the code, test it and deploy it. I make the decisions, I hold the keys, and I check the work before it counts.

See the applicationsWrite to me
  1. 1Specification
  2. 2Pre-flight
  3. 3Splitting
  4. 4Workbench
  5. 5Test bench
  6. 6Supervision
  7. 7Staging showcase
  8. 8Human verification station
The eight stations of the line, in conveyor order. The tickets moving along carry real ticket numbers.

As of 10 October 2026

  • 1,457

    code changes built by agents and accepted

    A "PR" (pull request) is a proposed change that is tested, then merged into the software.

  • 5

    applications built in 12 weeks

  • 7.5/10

    the agents' own average self-reported confidence, and only 3 scores of 10/10

  • 1

    human to decide and verify

How these figures are counted

Self-reported confidence scores

The 1,013 agent PRs that state a "Confidence: N/10" score, grouped by score. Average: 7.5/10.
  1. 40/10: 4 scores
  2. 01/10: 0 scores
  3. 02/10: 0 scores
  4. 23/10: 2 scores
  5. 24/10: 2 scores
  6. 135/10: 13 scores
  7. 1006/10: 100 scores
  8. 2787/10: 278 scores
  9. 5238/10: 523 scores
  10. 889/10: 88 scores
  11. 310/10: 3 scores

121 scores of 6/10 or lower (12%)3 scores of 10/10

The journey of a real ticket

Nine hours in the life of a ticket

Here is a real Merkindium ticket from 19 September 2026, step by step. Nothing is staged: it got stuck, it was picked up again, and the agent said what it had not checked.

The production line, in eight stationsEight stations arranged in a zigzag along a conveyor, from the specification (1) to the human verification station (8). Ticket #940 is born at the workbench (4), goes back to the specification (1) for the decision, is released at splitting (3), returns to the workbench, becomes PR #946, goes through the test bench (5), supervision (6), the staging showcase (7), then the human verification station (8). Without animation, the ticket rests at station 8.1Specification2Pre-flight3Splitting4Workbench5Test bench6Supervision7Staging showcaseVerified8Human verification station#940PR #946
This ticket did not go through pre-flight: it was born along the way.

This ticket did not go through pre-flight: it was born along the way.

  1. Step 1 ·

    The finding

    While working on another ticket, an agent notices that the forum's "New topic" screen does not match its mockup: two "Write / Preview" tabs where the mockup shows a single "Preview" button. It does not decide on its own: it opens ticket #940.

    #940opened at the workbench, during ticket #907

    The mockup

    a single "Preview" button

    The "New topic" screen

    two "Write / Preview" tabs

    Station 4 · Workbench, during ticket #907

  2. Step 2

    The human decision

    I decide: the mockup is the reference, the code gets aligned. The decision is written at the top of the ticket, with what to do and how to check it. The decision also goes into the specification, as question Q-137.

    "The mockup is the reference, the code gets aligned."At the top of ticket #940
    SpecificationQuestion Q-137

    Station 1 · Specification

  3. Step 3 ·

    The project-manager agent releases the ticket

    An agent schedules the work, in dependency order, with a limited number of tickets in progress at once.

    #940ready

    · "ready" label set by the project-manager agent, in dependency order

    Station 3 · Splitting

  4. Step 4 ·

    Building it, in five attempts

    On the first try the agent gets stuck after four minutes. The factory hands the ticket to another model and starts again. Four retries later, the work is ready.

    1. 1stuck at 4 min
    2. 2another model
    3. 3retry
    4. 4retry
    5. 5ready

    five attempts

    Station 4 · Workbench

  5. Step 5 ·

    The proposal

    The agent opens PR #946: one button instead of two tabs, typed text never lost, labels translated into French, English and Spanish, and tests to prove it.

    #940PR #946

    • one button instead of two tabs
    • typed text never lost
    • labels translated into French, English and Spanish
    • tests to prove it

    From station 4 to station 5 · the ticket becomes "PR #946"

  6. Step 6

    Automated tests

    Seven checks run. One of them, required by the ticket, could not run: another test was already failing on the main version. The agent says so in plain words instead of ignoring it.

    1. 1: ran
    2. 2: ran
    3. 3: ran
    4. 4: ran
    5. 5: ran
    6. 6: ran
    7. 7: could not run

    6 checks ran1 could not run, and the PR says so

    Station 5 · Test bench

  7. Step 7 ·

    The merge

    Supervision accepts the PR, the ticket closes, and the new version deploys itself to the staging server.

    "Staging" is a private copy of the application where things are checked before the public sees them.

    PR #946Merged · 15:04
    Ticket #940 closedStaging server

    Station 6, then station 7 · Supervision, then staging showcase

  8. Step 8

    Human verification

    The PR lists five steps to check it, and ends with an honest score:

    Recette · Merkindium · PR #946Confidence 6/10
    "Confidence: 6/10 — […] no real browser check."
    VerifiedBugSkip

    This is exactly why a human goes over the work.

    Station 8 · Human verification station

What the factory has built

Five applications, in the state they are really in

Each one keeps its own visual identity. The badge says where it stands as of 10 October 2026, without rounding up.

  • Liveactually in use.
  • In stagingcomplete in a private verification copy, not open to the public yet.
  • In progressstill being built.

Chartrium

In progress

Chartrium, "To check" queue: a fictional document read, the suggested folder and the "File here" buttonChartrium, a fictional document open with its record and versionsChartrium, "Who did what, and when" history
Real screenshots on fictional data: to check, document, history

Find a document fast, never lose a version, and be able to show what happened to it.

For whom
Small law firms (4 to 10 lawyers)
Stack
  • Go
  • React + TypeScript
  • PostgreSQL
  • Temporal
  • Apache Tika

Read the case study Chartrium

Merkindium

In staging

Merkindium, home page in light themeMerkindium, home page in dark themeMerkindium, tutorial catalogue sorted by categoryMerkindium, a tutorial pageMerkindium, a forum threadMerkindium, project gallery
Real screenshots: home, catalogue, tutorial, forum, gallery

The companion site of a home-DIY YouTube channel: sorted tutorials, courses, forum and gallery.

For whom
Viewers of Merkindium, in French, English and Spanish
Stack
  • Next.js 15
  • PostgreSQL + Prisma
  • Stripe
  • Tailwind

Read the case study Merkindium

Recette

Live (internal use)

Recette, verification queue: ticket #940 scored 6/10, "to watch", with the Verified, Skip and Bug buttons
Real screenshot: ticket #940 in the queue, scored 6/10

Check every factory delivery in three moves: Verified, Bug, Skip.

For whom
Me, the reviewer
Stack
  • Next.js
  • SQLite

Read the case study Recette

Bottrading

In progress · private access

Bottrading, design board: creating an agentBottrading, design board: interface statesBottrading, design board: colour token chart
Design boards: creating an agent, states, chart

Design, test and monitor trading agents on your own account, without the platform ever holding funds.

For whom
A private, invite-only circle
Stack
  • Next.js
  • Node engine

No performance, no returns, no investment advice: this shows the design, not gains.

Read the case study Bottrading

Raccourci

Live

Raccourci, home page: a field for the long address, a field for a custom code, and the Shorten button
Real screenshot: home page

The very first application: a link shortener used to tune the factory.

For whom
Personal use
Stack
  • Node.js
  • Fastify
  • SQLite

All applications

What stays human

Deciding.

What gets built, and what to do when the specification contradicts itself.

Holding the keys.

No password or payment key ever goes into the code: the factory builds without knowing them.

Verifying.

Every delivery goes through a human check before it counts.

How it works, in detail

Limits

  • Quotas.

    Subscriptions have caps. When they are reached, the factory slows down or stops.

  • What agents miss.

    They rarely check in a real browser. 121 of their confidence scores are 6/10 or lower, and they say so.

  • Flaky tests.

    A real share of the work goes into fixing tests that fail at random: 40 PRs out of 440 in Merkindium.

  • Getting stuck.

    A ticket can stall several times: the one on the home page took five attempts.

  • The human remains the bottleneck.

    Deciding, checking, unblocking: that time cannot be delegated.