What it costs (order of magnitude)
about €300 a month
servers and agent subscriptions included.
No usage-based cost: agents run on fixed-price subscriptions.
How it works
An eight-step chain. Agents at every station, a human at the three places that matter.
The production chain
Scroll down: a ticket moves from station to station, and each station comes alive when you reach its text.
Step 1 of 8
The specification
Requirements written so they can be tested, acceptance criteria, architecture decisions, mockups.
Station 1 · Specification
Step 2 of 8
Adversarial pre-flight
Three agents hunt for contradictions in the specification before anything starts. 59 findings on Merkindium, all resolved before the first ticket.
Station 2 · Pre-flight
Step 3 of 8
Splitting the work
The project-manager agent splits it into batches and tickets, with their dependencies.
Station 3 · Splitting
Step 4 of 8
Building
An agent takes a ticket and proposes a change (PR).
Station 4 · Workbench
Step 5 of 8
Testing
Every PR runs the automated tests, plus a real start of the application in a container.
Station 5 · Test bench
Step 6 of 8
Supervision
If everything is green, the PR is merged. Otherwise a fixer agent takes it back.
Station 6 · Supervision
Step 7 of 8
Staging
The new version deploys itself to a private copy.
Station 7 · Staging showcase
Step 8 of 8
Human verification
In Recette: Verified, Bug or Skip. A bug becomes a new ticket.
Station 8 · Human verification station
What the human does
Three things that stay with the human.
Decides
the specification, the trade-offs when it contradicts itself, what goes to production.
Holds the secrets
payment keys, passwords, server access. Agents see none of them.
Verifies
in staging, starting with what the agents doubt.
Safeguards
Five protections built into the chain.
A mandatory check (
ci-ok)Nothing is merged without green tests.
A daily budget per application
Each one has a daily consumption cap.
No secrets in the code
Checked automatically.
A stub mode for every external service
The factory builds without holding any key.
A limited number of tickets in progress
In dependency order.
The means
GitHub for tickets and tests, two families of agents (Claude Code and Codex), twelve build machines on servers rented from OVH, Coolify for deployments.
Limits
Quotas.
Subscriptions have caps. When they are reached, the factory slows down or stops, and the project-manager agent alone uses a large share.
What agents miss.
They rarely check in a real browser. They can ship code that passes the tests and is still wrong on screen. 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. When the main version breaks, the factory can stay stuck until I decide.
The human remains the bottleneck.
Deciding, checking, unblocking: that time cannot be delegated.
The figures, and how they are counted
Collected from the GitHub API on 10 October 2026; regenerated by script.
- 1,457
agent PRs, out of 1,486 merged PRs across the 5 applications shown
A "PR" (pull request) is a proposed change that is tested, then merged into the software.
- 1,681
closed issues (tickets, bugs, dropped ones included)
- 7.5/10
average of the "Confidence: N/10" scores the agents report about themselves
- 1,013
agent PRs that state a confidence score
- 121
scores of 6/10 or lower (12%), and only 3 scores of 10/10
- 12 weeks
from 19 July 2026 to 10 October 2026; first repository on 19 July 2026
Self-reported confidence scores
121 scores of 6/10 or lower (12%)3 scores of 10/10
| Application | Agent PRs / merged PRs | Closed issues | Average confidence |
|---|---|---|---|
| Chartrium | 621 / 640 | 720 | 7.5 |
| Merkindium | 440 / 445 | 525 | 7.4 |
| Bottrading | 201 / 202 | 216 | 7.7 |
| Recette | 179 / 183 | 194 | 7.6 |
| Raccourci | 16 / 16 | 26 | no score |
| Total | 1,457 / 1,486 | 1,681 | 7.5 |
A PR counts as "built by an agent" when it comes from a ticket branch (ticket/…). All of them are opened under my GitHub account, which the factory uses. The confidence score is self-reported by the agent. The rule came in along the way: Raccourci did not have it, nor did some older Chartrium and Merkindium PRs. Collected from the GitHub API on 10 October 2026; regenerated by script.
5 applications shown.