Case study · Our own work
Built in days, not months, with one person directing.
Our own platform, mibble, is built every day by a software factory: AI agents write the changes, and the pipeline decides what needs a person.
AI and software builds · 4 min read01The leak
The person approving was the bottleneck.
Before the factory we shipped about 150,000 lines of code a month. Even in September, a person still approved 70% of changes, and the pipeline waited on them.
02The fix
Checks first, people for what matters.
A handful of rules, enforced by the tooling rather than written in a document.
- Checks first automatic tests and checks run on every change, however small.
- Approval by risk only changes that are hard to undo wait for one person.
- Nothing clears its own work the agent that wrote a change can never be the one that clears it.
Lines of code shipped each monthBeforeWith the factory
4× Oct vs the Jun–Sep average of 482k a month
03Now
Over 2 million lines in October.
In October the factory shipped 1,298 changes and over 2 million lines of code. Over the 30 days to 6 November, 91% were cleared by the automatic checks alone. The same discipline sits behind every AI and software build we do for clients.
From our own repositories: code lines changed (code only, no generated files, lockfiles, data or docs) and merged changes by month, March to October 2026.
The detail, in our field notes
- How We Run Our Software Factory (and What Broke)
- We Were the Bottleneck: What 3 Days of Data Taught Us About Our Own Automation
- Stop Measuring AI by Test-Pass Rate
- Our Journey to Automated Code Review (We're Still Figuring It Out)
- Is Jujutsu a Git Superpower for AI Coding?
- To Branch or Not to Branch: Why We Build on Trunk With AI Agents
- Why GitHub Wasn't Built for AI Agents (and What to Do About It)
Got a leak like this?
Tell us what keeps slipping. A fixed-scope growth review finds it, measured in sales and margin.