What is a software factory?
A software factory is a repeatable system for producing software the way a factory produces goods, with standard inputs, automated production, quality checks and a feedback loop. The term's four meanings since 1969, how an AI software factory works in 2026, and how it scales down to a personal software factory.
The short answer
A software factory is a repeatable system for producing software the way a factory produces goods. Work goes in as a defined specification. Standard tools, reusable parts and automation produce the code. Quality checks decide what is allowed to ship. A feedback loop improves the process for the next run. The word "factory" refers to repeatability, not size. The point is that the tenth application, or the tenth change, comes out as reliably as the first.
The term has been in use since 1969 and its meaning has shifted with the technology. In 2026 it most often means an AI software factory: a pipeline in which AI coding agents do most of the implementation while engineers write the specifications, set the checks and decide what ships. A personal software factory is the same idea scaled down to one person or a small team. This site, Personal Software Factory, publishes specs and test results for that scale.
The four meanings of "software factory"
The same phrase has meant four different things. They share one idea: take the knowledge from building similar software and put it into a process instead of leaving it in individual people's heads.
| Era | What "software factory" meant | Who used it | What the factory standardised |
|---|---|---|---|
| 1969 to 1990s | A large, disciplined software organisation run on manufacturing principles | Hitachi (Hitachi Software Works, 1969), System Development Corporation (1975), then NEC, Toshiba and Fujitsu | Process, tools, training, reuse of code and designs |
| 2004 to 2010s | A software product line: a package of models, templates, frameworks and guidance for building one kind of application | Microsoft patterns & practices, after Greenfield and Short's book Software Factories (2004) | Architecture, code generation, templates |
| 2017 to today | A DevSecOps organisation that ships government and enterprise software continuously | US Air Force (Kessel Run), US Army Software Factory, VMware Tanzu Labs clients | CI/CD pipelines, platforms, team practices |
| 2025 to today | An AI software factory: agents build, engineers decide what ships | Factory ("Factory 2.0", June 2026), LaunchDarkly, CircleCI, Warp | Specs, agent execution, automated verification, feedback |
1. Japan's software factories (1969 onwards)
Hitachi was the first company to call a software organisation a factory, with Hitachi Software Works in 1969. System Development Corporation in the US followed in 1975, described by Bratman and Court in the IEEE journal Computer as "The Software Factory", and NEC, Toshiba and Fujitsu set up their own in 1976 and 1977. MIT's Michael Cusumano studied them in Japan's Software Factories (1991). These were not automated: they were large groups of programmers working to standard methods, with shared tools and a deliberate effort to reuse designs and code across projects.
2. Microsoft's Software Factories (2004)
Jack Greenfield and Keith Short, with Steve Cook and Stuart Kent, published Software Factories: Assembling Applications with Patterns, Models, Frameworks, and Tools in 2004. In their sense a software factory is a software product line: a schema, reference implementation, templates, code generators and how-to guidance for producing variants of one kind of application. Microsoft's patterns & practices group shipped several, such as the Web Client Software Factory. This is the definition Wikipedia still leads with.
3. DevSecOps software factories (2017 onwards)
From the late 2010s the US military used "software factory" for in-house teams that build and ship software continuously instead of buying it through multi-year contracts. Kessel Run, started by the US Air Force in 2017, is the best-known example. The US Army Software Factory trains soldiers and civilians in modern software development. VMware's definition from this period: "an organized approach to software development that provides software design and development teams a repeatable, well-defined path to create and update software." Here the factory is mostly the platform and pipeline: source control, CI/CD, security scanning and deployment, run as a shared service.
4. AI software factories (2025 onwards)
Once AI agents could write and test code, "software factory" came to mean a pipeline in which agents do the production work. Factory, the company behind the Droid coding agents, described it in June 2026 as a system that "starts with signals from the outside world: bug reports, internal conversations, customer feedback, business requirements", turns them into planned changes, and builds, tests, reviews, secures, ships and monitors them in a continuous loop. LaunchDarkly's October 2026 definition is a "repeatable development system where agents do most of the implementation and engineers decide what ships."
How an AI software factory works
Every version has the same parts as a physical production line. In an AI software factory they look like this:
| Part of the line | What it is | What goes wrong without it |
|---|---|---|
| Orders | Bug reports, feature requests and business needs, written up as a specification the agents can read | Agents build what is easy to guess, not what was needed |
| Design | A spec: data model, business rules with examples, screens, edge cases, acceptance criteria | Every build makes different assumptions |
| Machines | AI coding agents (Claude, Codex, Cursor, Factory's Droids and others) working in isolated environments | Humans become the bottleneck for every line of code |
| Production plan | The work split into steps small enough to check one at a time | Large changes that nobody can review |
| Quality control | Automated tests, type checks, security scans, acceptance checks run in a browser, human review of what ships | Code that looks right and behaves wrong |
| Feedback loop | Failures, gaps and production signals turned into fixes to the spec, the tests and the agents' instructions | The same mistake on every run |
The factory, not any one agent, is the product. Running ten agents at once is easy; getting ten changes that are correct and can be merged is the hard part. The quality-control and feedback stages are what separate a factory from a fast coding assistant.
Software factory vs. related ideas
| Term | How it relates to a software factory |
|---|---|
| DevOps | A set of practices and a culture: development and operations share responsibility for delivery. A software factory is a concrete system that puts those practices into an automated line. |
| CI/CD pipeline | One part of the factory: the stage that builds, tests and deploys a change. A factory also covers intake, specification and feedback. |
| Platform engineering | The team and internal platform that often run the factory for an organisation. |
| Software product line | The 2004 meaning: a factory for variants of one kind of application. |
| AI coding assistant | One machine on the line. It helps with a single task; a factory moves every task through the same steps and checks. |
| AI factory (NVIDIA's term) | Different thing: the compute infrastructure used to train and run AI models, not a way of producing software. |
| Dark factory | A fully automated factory that runs without people on the floor. Applied to software, it means a line where agents ship without human review; most teams keep a person deciding what ships. |
Enterprise software factory vs. personal software factory
Most writing about software factories is about large engineering organisations: many agents, many repositories, model routing, security review and governance. The same structure works at a much smaller scale.
| Enterprise software factory | Personal software factory | |
|---|---|---|
| Who runs it | A platform or engineering team | One person or a small team |
| What it produces | Changes to large, existing products | Small tools built from scratch, often internal tools |
| Orders | Tickets, incidents, customer feedback | One sentence from the person with the problem |
| Design | Specs written per change | One spec per tool, reused and improved |
| Machines | Fleets of coding agents, CI/CD, review bots | One AI app builder, such as Claude, ChatGPT, Cursor, Lovable, Replit or Type |
| Quality control | Test suites, security gates, staged rollouts | Acceptance checks, clicked through as named users |
| Feedback loop | Monitoring and incident data | Every place the builder had to guess becomes a fix to the spec |
A personal software factory is how a team without engineers gets software built to its own rules instead of paying per seat for a SaaS product: an approval workflow, a simple CRM, a time-off tracker. What is a personal software factory? covers it in full.
What we learned running a small software factory
Personal Software Factory (personalsoftwarefactory.com) runs a small software factory in public. We write a spec for an internal tool, an AI builder follows it phase by phase, and every acceptance check is run in a real browser. Across 20 build tests the builders exposed 386 gaps in our own specs. Three lessons from that work apply to software factories of any size:
- The spec is the bottleneck, not the code. The gaps were in the spec: sections that contradicted each other (41), fields the rules depended on but never defined (49), undefined edge cases (38) and ambiguous wording (34), among others. One example: what a help desk should do when a ticket is reopened after its agent has left. Where the spec was silent, the builder guessed, and the guess was often plausible and wrong. What AI app builders get wrong lists the gaps by type.
- Checks must test behaviour, not appearance. A screen can render correctly and still save the wrong value. Acceptance checks that click through the app as a named person, with seed data built so that a wrong rule fails, caught what a glance would not.
- The feedback loop is what makes it a factory. Each gap was fixed in the spec, so the next build from the same spec starts from a better design. Without that loop, an AI builder is a fast way to produce a new set of guesses every time.
The blueprint library has the specs, and How we test explains the method.
Frequently asked questions
Who coined the term "software factory"?
Hitachi was the first company to use it for an organisation, with Hitachi Software Works in 1969. Bratman and Court's 1975 paper "The Software Factory" described System Development Corporation's version, and Greenfield and Short's 2004 book gave the term its software-product-line meaning.
What is an AI software factory?
An AI software factory is a software factory in which AI coding agents do most of the implementation. Engineers write the specs, set the checks code must pass and decide what reaches users. Work moves through the same stages every time: intake, specification, agent build, automated verification, review, release and feedback.
What is a software factory in the military?
In the US Department of Defense, a software factory is an in-house team and platform that builds and ships software continuously with DevSecOps practices, instead of buying it through long acquisition contracts. Kessel Run (US Air Force, 2017) and the US Army Software Factory are well-known examples.
Is a software factory the same as DevOps?
No. DevOps is a set of practices and a culture. A software factory is a specific, repeatable production system that uses those practices, and an AI software factory adds agents that do the building.
Do I need an engineering team to run a software factory?
Not at small scale. A personal software factory needs a precise spec, an AI app builder and the discipline to run acceptance checks and fix the spec when something fails. It suits internal tools used by a small group; regulated software and anything handling payments or medical data still needs engineers.
Where does the name Personal Software Factory come from?
From this idea scaled down: a software factory one person or a small team can run. See What is a personal software factory? and About.
Get new blueprints by email
One email when a new blueprint is published. Nothing else.