Lab · Minds · A series · Interactive, in your browser
Software factories: an old word for a new line
“Software factory” is an old phrase, and it means something new now. This series covers what it has meant since Bemer used it in 1968, how far a person can step back from the code, what stops an agent cheating its way past a test and where a lesson should live. Then, what has been measured.
The word is old
Bemer’s working paper at the 1968 NATO conference on software engineering, in Garmisch, called a machine-controlled production environment a “software factory”. Program construction, checkout and use: one set of tools for all three.
Bemer’s aim, as the report quotes it: “Ideally it should be impossible to produce programs exterior to this environment.”
McIlroy, at the same conference, argued that the field had no components subindustry, and floated a pilot plant for components. I use his heading for the idea: “A Components Factory”.
In 2004 Greenfield described Software Factories, a method Microsoft developed, as a development environment set up to build one type of application quickly. He was after economies of scope, many similar products. Not economies of scale.
Development, Greenfield wrote, would never become a purely mechanical process. The point was letting developers encapsulate what they know, as reusable assets.
These pages don’t tell all the history between 1968 and 2004. I couldn’t read the sources I wanted for that stretch, so I left it out. Better than guessing.
The two versions I could read, Bemer’s and Greenfield’s, industrialised the people doing the work with better tools, reusable parts and a disciplined process. The 2026 version swaps out who does it.
The word in 2026
In February 2026 StrongDM’s AI team published “Software Factories And The Agentic Moment”, which is a vendor describing its own practice. It sets two rules. Humans must not write the code, and humans must not review it.
The last of Shapiro’s five levels is a “dark factory”, named after lights-out manufacturing: the person isn’t part of the process. The framework is his, built from what he’s seen.
No two factories are the same, says Jacob, who founded Swamp Club, because each depends on the application and the team. The software it should produce sets its needs. Everything else should adapt to them.
Start small, Jacob’s post of 5 November 2026 says, then encode more of what makes your software yours. In his words: “It’s the time for discovering the small primitives and shifts in thinking”.
The promise is the same as ever: a codified process that turns intent into working software, again and again. People design the line now, meaning what it’s told, what it has to pass and where a person signs off. Agents write the software.
Three eras, side by side
1968
Bemer, NATO conference
- Who does the production work
- People programming inside one environment, with its tools.
- What is standardised
- The environment: computing power, a file system, compilation, and building the test and final systems.
- Where a person signs off
- Not stated
- What is measured
- Not stated
2004
Greenfield, Microsoft
- Who does the production work
- Developers, with the rote work automated. It will never be a purely mechanical process, Greenfield says.
- What is standardised
- Reusable assets and patterns for one type of application.
- Where a person signs off
- Not stated
- What is measured
- Not stated
2026
StrongDM, Shapiro, Jacob
- Who does the production work
- Agents. People don’t write the code or review it, by StrongDM’s rule.
- What is standardised
- What each team encodes for its own software, by Jacob’s account.
- Where a person signs off
- It varies. At Shapiro’s level 3 a person reviews code, and at his level 5 they’re out of the process.
- What is measured
- No outcome metrics on StrongDM’s page. What has been measured gets weighed in the fourth chapter.
The parts, and where they are on this site
A factory has these parts, and here’s where each one is covered. Where a linked page already explains something, this series doesn’t do it again. The word harness and how my own loop is built are on Harness.
- Agents working in parallel, and why more of them does not mean fasterThe mythical man-month
- The shapes of a line: router, workers, evaluator, gates and budgetsThe agent graph
- Records of a run, and proof that a check passed on this contentTraces and receipts
- How far a person steps backC1, How dark is the factory?
- Checks an agent cannot gameC2, Tests can be gamed
- Where a lesson livesC3, Encoding what makes your software yours
- What has been measuredC4, What the factories have shown
- A person’s attention as the scarce thingAttention is the bottleneck
The chapters
C1 · Open
How dark is the factory?
At each level of autonomy, where does the person stand, and what has to exist before they can step back?
C2 · Open
Tests can be gamed
When agents write the code and can touch the tests, what stops them passing by cheating?
Builds on C1 How dark is the factory?
C3 · Open
Encoding what makes your software yours
Where should a lesson live so the next run starts with it, and which of those places actually enforce it?
Builds on C2 Tests can be gamed
C4 · Open
What the factories have shown
What has been measured about factory-style work, as opposed to claimed?
What this series is not
- A buying guide. I don’t recommend a vendor or a tool, and nothing here is for sale.
- A measurement. No factory runs on this site. The examples in the figures are invented.
- A cost comparison. If a source reports a cost, I say so and leave out the amount.
Limits
Bemer’s 1968 paper and Greenfield’s 2004 article describe a plan and a method, not results.
StrongDM’s page is a vendor’s, and Shapiro’s and Jacob’s posts are opinions. None of them is a measurement. That gets weighed in the fourth chapter.
Sources
Bemer, 1968 (paper) Machine-controlled production environment, in the report of the NATO conference on software engineering. Report on a conference sponsored by the NATO Science Committee, Garmisch, October 1968, section 5.3.1, page 55, from Brian Randell’s copy; read 2026-11-05.
The report calls the working paper the most ambitious plan for production tools at the conference. It names the machine-controlled production environment, or software factory, where program construction, checkout and use happen entirely inside one environment.
- Section 5.3.1, the machine-controlled production environment, or software factory
- Section 5.3.1, report page 55
- Section 5.3.1
McIlroy, 1968 (paper) Mass Produced Software Components, in the same report. Report on a conference sponsored by the NATO Science Committee, section 8.2, with the heading “A Components Factory” on page 84; read 2026-11-05.
The absence of a software components subindustry as one weakness of the field, and a pilot plant for components. His essay never uses the phrase software factory.
- The opening and the section headed A Components Factory, report page 84
Greenfield, Microsoft, 2004 (vendor) The Case for Software Factories. Microsoft’s own article in The Architecture Journal, July 2004, in Microsoft’s archive; read 2026-11-05.
A Software Factory as a development environment configured for one type of application; economies of scope, not scale; development never purely mechanical; letting developers encapsulate what they know as reusable assets.
- Summary; Economies of Scale and Scope
- What Will Industrialization Look Like?
- Summary; Raising the Level of Abstraction
- Summary
StrongDM AI, 2026 (vendor) Software Factories And The Agentic Moment. The vendor’s own page, 6 February 2026; read 2026-11-05.
The two rules (code must not be written or reviewed by humans), scenarios kept outside the codebase like a holdout set, satisfaction in place of a green suite, behavioural clones of outside services, a daily token-spend benchmark per engineer with no amount shown here, and no outcome metrics.
- The two rules
- The page gives none
- The page
Shapiro, 2026 (opinion) The Five Levels: from Spicy Autocomplete to the Dark Factory. Blog post, 23 January 2026; read 2026-11-05.
Levels 0 to 5 as he describes them, where he places himself, why he calls level 5 dark, and his own estimates of who is where, which are observation and not data.
- Level 5 and Why Dark?
- Levels 3 and 5
- The post
Syntax episode 1049, 2026 (founder interview, podcast) WTF is a Software Factory? (Adam Jacob with Wes Bos and Scott Tolinski). Show page and transcript, 21 October 2026, timestamps from the transcript; read 2026-11-05.
No two factories are the same; black-box acceptance tests first; doubts about how much unit tests add now; the architecture written down as a skill; changing the factory when a failure gets through; skills inside a system with inputs, outputs and measurement.
- 10:44 to 13:48, a factory takes most of the development loop and none is the same
- 10:44 to 13:48
Jacob, 2026 (founder’s opinion) Adaptive Building Blocks. Blog post, 9 April 2026; read 2026-11-05.
His factory has its own needs, set by the software it should produce, and everything else should adapt to them; Swamp’s development moved off GitHub as more of the lifecycle happens inside the agent.
- His factory has its own needs
Jacob on X, 2026 (founder’s opinion) Post on X, 5 November 2026. Post on X, read in Claude in Chrome (x.com refuses plain fetches); read 2026-11-05.
Start small, build the basic thing, then encode more of what makes your software unique; small primitives over big answers.
- The post
Each fact is paraphrased from the page named, read on the date shown. A company’s own page is marked as the vendor’s claim, an essay or a post as opinion, and the survey as vendor-run. None of it is a measurement of this site. Where a sentence is my own reading, it is marked “My view”.