From writing code to directing factories.
A software factory is a production line, not a project. Specifications go in, agents build in parallel, evaluation gates every change and a product comes out. One person directs several lines, one per business. We design the line, run it with you and hand you the controls.
Software factories
Agent production lines: specs in, agents build, evals gate, product out.
What we do
Specification first
We turn your idea into specs an agent can execute: scope, acceptance criteria and the tests that define done.
Agents that build
Claude Code, Pi or Codex working in isolated environments, in parallel, loaded with the skills of your stack.
Evaluation that gates
Tests, a review by a second agent, security checks and a human on the merge to production. Nothing ships on trust.
Direction, not execution
A board where you see every line, its cost and its output. You decide what to build next; the agents execute.
How it works
- 1
Pick one line
One product or one business with a backlog bigger than the team. We start there, not everywhere.
- 2
Write the first specs together
Three to five specs that cover a real slice of work, with their acceptance tests.
- 3
Run the line for a sprint
Agents build, evals gate, you direct. We tune the skills and the checks until the output is boring.
- 4
Add lines
A second product, a second business. Same factory, new specs, one seat of direction.
What you get
- Factory setup: repositories, CI, agents, skills and environments
- Spec templates and acceptance-test conventions
- Direction board with cost and output per line
- A playbook for directing several lines
Good fit if
- A founder or team with several products, or many small ones
- A backlog larger than the people available
- Willingness to write specs instead of tickets
Not a fit if
- A single heavily regulated product with no tests yet (start with Build tools)
- Teams that want agents to decide what to build
Stack we use
Questions
Does anyone review the code?
Yes. Evals and a second agent review every change, and a person approves merges to production.
How do you keep costs in check?
Each line has a token budget and a cost line on the board, with the same observability we use in Cut AI costs.
Is this how you work yourselves?
It is. Our own office of AI employees, starting with a content creator on a kanban, runs on this model.