The State of Software Factories in 2026

Last week, Alan Wells posted on X about seven days of defense against an active cyber attack on one of his products. His team has three people. They survived, he wrote, because they had spent the months before building a software factory. His conclusion: “Software factories are going to be required equipment for everyone.”

We run something like a software factory at Astrohacker. Agents solve our issues with experiments, a command called ah keeps the records, and we choose for each issue who reviews the work. So I wanted to know what other people mean by the term, who runs one, and how. This post is what I found.

One warning first. Most of the sources are vendor blogs, conference recaps and secondary summaries. Most company numbers are self-reported, and some of them disagree with each other. I give both numbers where they disagree.

What a software factory is

The term is old. It goes back to a 1968 paper by Bob Bemer. In 2026 it has a new meaning: a system where AI agents take work from a queue, write the code, test it and open a pull request, while people decide what to build and what ships.

It is not an “AI factory”. That term is NVIDIA’s name for data centers that train and serve models.

Addy Osmani gives the clearest outline, in Software Factories, Light and Dark:

• A loop is one agent that does one job again and again until a check passes.
• A harness is the environment around a loop: its sandbox, tools, memory, and the gates that define “done”.
• A factory is many harnessed loops that take work from a queue, pass through review gates and are supervised by people.

Warp lists the stages of the loop as triage, spec, implementation, review, verification, ship and monitor. Factory closes the loop: monitoring produces new signals, such as bugs and feedback, and the signals become new work.

Light factories and dark factories

The most important difference between factories is how much people review.

• A light factory: agents do most of the building. People review the work where a wrong decision is expensive, and a person merges each pull request.
• A dark factory: agents write, verify and ship code that no person reads. The name comes from manufacturing plants that run with the lights off.

Almost everybody runs a light factory. One analysis found that every operator and vendor that it examined delivers agent work as a pull request for a person to read.

The best-known dark factory is at StrongDM. In February, its AI team described two rules: “Code must not be written by humans” and “Code must not be reviewed by humans.” Specs and test scenarios drive the agents. The agents prove each change against a “Digital Twin Universe” of fake copies of services such as Okta and Slack. The CTO’s measure of a good factory is at least $1,000 of tokens for each human engineer each day.

The dark factory is contested. Simon Willison asked how it could be sensible, because LLMs make “inhuman mistakes”. Stanford’s CodeX asked who is responsible when nobody reads the code, especially at a security company. And HumanLayer ran a dark factory from July 2025 for some months. They stopped after an outage forced weeks of work through code that nobody had ever read.

Osmani has a name for that risk: comprehension debt. It is the gap between how much code exists and how much of it people understand. A dark factory makes this debt quietly, and the bill comes late.

Who runs one

These are the factories with public records:

• Stripe: Minions. An engineer starts one from Slack. Reports say the Minions merge more than 1,000 pull requests a week. Agents write all of the code, and people review it. Each agent runs in a “Devbox”, the same warm EC2 sandbox that human engineers use, which starts in about ten seconds. About 500 internal tools are available to the agents through MCP.
• Ramp: Inspect. Each session gets its own Modal sandbox with a full development environment, the OpenCode agent and a browser for visual checks. Snapshots of every repository are rebuilt every 30 minutes, so sessions start fast. Reports say Inspect writes from 30% to more than half of Ramp’s merged pull requests.
• Uber. In August, Uber wrote about running a software factory at Uber scale. More than 70% of its pull requests come from local or cloud agents, and its engineers have written more than 3,600 agent skills. Agents also review code, repair failed CI runs, triage on-call alerts and debug bugs.
• OpenAI. In February, a team of three engineers, later seven, built a product of about one million lines in five months with no code written by hand.
• Others. Cloudflare says 93% of its R&D is now “delegated engineering”. Spotify reports 650 agent pull requests a month. Vercel reports that agents write a quarter to a third of the pull requests for its AI SDK. Warp says agents make 20–30% of its own pull requests.

The factory is not only for large companies. Alan Wells’s team has three people. Theirs uses an opencode server, a fork of compound engineering, and an agent in Slack that runs scheduled jobs.

The bottleneck is verification

Every source agrees on one point: generating code is the easy part. Verifying it is the hard part.

Stripe’s Minions work because of what Stripe built before agents: the devboxes, a suite of 3 million tests, a lint daemon and the CI pipeline. Alan Wells’s team stopped new product work for about six weeks to build validation good enough to ship a pull request without human review. He wrote that it was painful, and worth it.

Osmani says autonomy can grow only as far as changes can be checked cheaply and reliably. Good checks are fast and hard to game: types, property tests, browser tests and review agents.

Where the agents run

Mostly in the cloud. At Ona’s Background Agents Summit in May, almost every company system ran in isolated cloud sandboxes. The teams had reached the same five parts on their own: sandboxed environments, connections to context, triggers, fleet orchestration and governance.

The cloud gives many agents in parallel, a fast start from warm snapshots, a small blast radius when an agent makes a mistake, and work that continues while the engineer sleeps. Local agents can use the whole machine and its hardware. Some products sell both: Cursor and Docker let you start on your laptop and finish in the cloud, and Factory offers hosted, self-hosted, air-gapped and local machines.

Wells gives the minimum: a shared coding server in the cloud with read-only access to all of your logs and monitoring. Then you are “a prompt away” from good knowledge of the situation. And because everyone can see each session, incident response becomes multiplayer.

Most factories stop at the pull request

The usual factory goes from an issue to a pull request, and then a person merges it. Deploy is a separate step.

Checkly wrote about this last week: Your software factory stops at deploy. A factory that ships a broken pay button learns about it from a customer ticket. Checkly’s answer is a synthetic browser check in production. When it fails, it starts an agent with the evidence, and the agent opens a fix. A few teams let agents merge or roll back after the checks pass. Most stop at a draft pull request.

So the full loop (an issue goes in, and a verified fix comes out in production with no person) is rare. It is the dark factory again, with deploy added.

The market

Nobody publishes a market size for a “software factory” as such. These are the closest numbers for 2026:

• AI coding agents for enterprises: about $10–11 billion a year (Gartner). Estimates for all AI coding tools are from about $9 billion to about $13 billion.
• Custom software development, the “build it for me” market that a factory can replace: about $51–65 billion, with growth of 17–22% a year.

There are company signals too. Cognition, which makes Devin, reported $492 million of annual recurring revenue in May, up from $37 million one year before. That number includes Windsurf. Factory was valued at $1.5 billion in April, and reports say $5 billion in September.

Two kinds of product are on sale:

• A factory you rent. Buildful calls itself “the software factory you can buy”: 1,000 tasks for $10 a month. It goes from a task to a pull request in a cloud sandbox, and a person merges.
• A factory built for you. BCG Platinion sells the design and pilot of an agentic software factory as consulting. Platforms from Factory, Warp and Ona sell the parts.

What this means for us

Our factory runs on a Mac, not in the cloud. It is a light factory with a dial: for each issue, we choose who reviews the design and the result of each experiment (a person, an agent, or nobody). And it has one thing that I did not find anywhere else: a written record for each experiment, with its decision, its evidence and its result, Pass or Fail.

We are now deciding how this changes Astrohacker. Zach Lloyd of Warp said it this month: software engineering is becoming factory engineering. If he is right, every team will need a factory, and most teams will need help building one.

If you run a software factory, or you want one, I want to hear from you. Get in touch.