La Forge

The problem: coding agents you have to watch

Coding agents can write code. Letting them work unsupervised on a real repository is another matter. Who checks what they did? What stops them running something dangerous? And how do you keep a record of every decision?

La Forge is our answer, and the tool we use for our own development work.

A production line, one ticket at a time

The idea is easy to describe. La Forge takes a repository and a stage of work (architecture, specification, implementation and so on), runs the chosen agent in a throwaway container, commits the result to a branch, opens a pull request and files a detailed report.

Each change is then reviewed by several different models, and their comments are merged into one. A human stays in charge of the sensitive stages: some gates only open with an explicit sign-off.

Guard rails instead of trust

Containers never see a production secret. A guard intercepts commands before they run and blocks anything that has no business there, such as reading environment variables, mass-deleting files or piping a downloaded script straight into a shell. A validator also checks the ticket backlog, its dependencies and any cycles before work even starts.

Everything is driven by a single configuration file: which models, which limits, which repositories, which stages need a person. Switching the main agent takes one command and leaves the rest untouched.

Why it matters to you

La Forge runs on our own server and keeps a full log of every run. For clients, that means faster development without giving up review or tests. And if you want to automate part of your own software production, we know the ground well: we built this, and we use it.