Case study · 04
Software Factory
The workflow we ship with: one AGENTS.md file and a handful of agent skills that keep Claude Code, Codex, and Cursor on the same process for every task.
1,200+
4
5/5
What we built
The Software Factory is how we build software with AI coding agents. It isn't a
platform. It's one AGENTS.md workflow file and a collection of agent skills,
each a folder with a SKILL.md that the agent loads when a task matches. Drop
them into a repo and Claude Code, Codex, and Cursor follow the same process on
every task.
The skills are open source on GitHub, and the full walkthrough is on YouTube, so any team can copy the setup.
How it works
Every task moves through the same four beats, each backed by a skill.
- Isolate
- Every task starts in a fresh Git worktree branched from
origin/main(new-feature), so several agents can work in one repo at once without conflicts. - Build
- Code follows a service-layer architecture (
code-structure): actions decide why and when, and a shared service layer owns the reusable how. - Prove
- The repo's checks plus runtime evidence (
evidence-driven-testing): the before state captured while reproducing the issue, and the after once the change works. - Ship
- The pull request opens with before/after proof in its description (
before-and-after), then review loops run until Greptile reports 5/5 with zero unresolved comments (greploop).
A short set of multi-agent rules holds it together: never commit to main,
one worktree and one branch per agent, a scope check against open pull
requests before starting, and stop and ask when a conflict can't be resolved
confidently.
Why it matters
Most teams adopting coding agents get speed and lose consistency: every agent, and every session, does things a little differently. A written workflow that every agent follows is how you keep the speed and still trust what ships. It's the same process we set up for client teams.