Skip to content

Examples of software factories

Examples of software factories range from one operator with a queue to companies whose agents account for most pull requests, with or without review.

Last updated , 8 min read

What are well-known examples of software factories?

Well-known examples of software factories include 5 whose builders have described them in public. StrongDM describes a factory with no human code review. Stripe, Uber, and Ramp describe systems that keep people in review or in the loop. One operator's published log describes a factory that one person runs.

A software factory is a repeatable process in which coding agents take work from a queue, build it, check it, and ship it. People decide what to build and try what comes out. In each example, agents build with no person at each step, and a check outside the builders' control comes before a change ships. Setting up the tools and checks around the agents is harness engineering.

The shape that five published software factories share Agents buildno person at each step A check before the change shipsthe examples differ here The change shipsmerged or deployed Stripelint and tests Uberreview agents Rampvisual validation One operatoragents verify each change StrongDMscenarios, often kept outside people review or stay in the loop a person tries each result no human review, by rule
The five examples share one shape and differ in the last check. The small line under each name is the automated check that its own source describes.

The examples differ in who reviews the code and what checks each change:

ExampleHuman review of the codeAutomated check the source describes
StrongDMNone, by ruleScenarios, often stored outside the codebase
StripeAn engineer reviews each pull requestLint and automated tests
UberHuman reviews and escalationsAgents that review code and validate changes
RampA pull request prepared for a person's reviewVisual validation
One operatorNone, but the operator tries each resultAgents that verify each change

Which factories run with no human code review?

Two examples on this list have no human code review. StrongDM states that as a rule, and the form has a name, the dark factory. The other is the factory of one described below, where a person still approves each goal and tries each result.

What does StrongDM report about its factory?

StrongDM's AI team describes its software factory as development in which specifications and scenarios drive agents that write code and run checks, with no human review. Its post states two rules, "Code must not be written by humans" and "Code must not be reviewed by humans."

A scenario is a user story that the software must satisfy from start to finish. The post says a test stored in the codebase can be rewritten to match the code. It also says that scenarios are often stored outside the codebase. A holdout test suite works on the same principle.

A large language model (LLM) can validate a scenario, a setup called LLM-as-a-judge. The team tracks the share of runs that would likely satisfy a user.

Which factories keep a person at review?

Three companies on this list describe people who review the work or stay in the loop. Stripe and Ramp do not use the word factory in the posts cited here. They are on the list because agents write the code and a person still checks the result.

What does Stripe report about its coding agents?

Stripe's engineering blog says that over a thousand pull requests merged each week at Stripe are produced by its coding agents. People review them, and they contain no code that a person wrote. The post calls the agents unattended, which makes them background coding agents.

A typical run starts from a message and ends in a pull request that passes continuous integration (CI). A test failure with no automatic fix goes back to the agent. The engineer who started the run then asks another engineer for review.

The agents work in Stripe's existing codebase with the same developer tools as its engineers. They also read the rule files that the engineers' coding agents read, files of the AGENTS.md kind.

What does Uber report about its software factory?

Uber's engineering blog says that more than 70% of its pull requests are attributed to local or cloud agents. The post uses the name software factory for the whole system. It lists the work that its managed agents handle, from code review to code maintenance tasks, and ends the list "with human reviews/escalations."

Most of the post is about cost. Uber's engineers split the cost of an agent session into separate terms and measure each one. They want the two usage terms to grow, and they work to reduce the rest. They also score their review agent against real pull requests with known bugs.

What does Ramp report about its coding agent?

Ramp says that by October 2026 its internal coding agent, Inspect, accounted for 74% of the company's merged pull requests. A person hands Inspect a task from a tool the team already uses, e.g. a chat message.

The post describes the result as a "visually validated and fully reviewed pull request." It also says that Inspect "prepares the pull request for your review." The post names the team's own judgment about what should ship as the bottleneck that remains. The same post opens early access to Inspect for other companies, and this entry covers only Ramp's own use.

What does a factory of one look like?

A software factory does not need a team around it. One published example has a single person in it.

What does one operator's log report about a factory of one?

In one operator's published log, a software factory deployed to production 72 times in 31 days. Its median change touched 4,578 lines of code across 47 files. The log describes the factory as a queue, agents that follow a set process, and one person.

The operator types short requests, answers the decisions that a goal leaves open, and tries what comes out. The operator does not read the code. Agents verify each change, and the log says that verifying and fixing the work takes as long as building it, on average. The log also lists costs, e.g. a queue that sat idle overnight with nothing to alert the operator.

What does this list leave out?

The list has four gaps:

  • The figures are the companies' own. Each number comes from the company or the person that runs the factory. None comes from an outside audit.
  • Only published accounts appear. A team that runs a factory and writes nothing about it is missing. So is a factory that a team tried and then dropped.
  • The examples are not ranked. The sources count different things, e.g. pull requests attributed to agents in one source and deploys in another. The order here follows the kind of review, not the results.
  • Products for sale are left out. A factory that a vendor sells to other teams as a product is a packaged software factory. Each example here is a factory that a team built for its own use.

How do these factories check what they ship?

Each factory on this list pairs its builders with a check that the builders do not control. The checks include scenarios kept outside the codebase, tests followed by a human reviewer, and agents that verify each change. Whatever shape your factory takes, give it one stage that runs the finished software and reports evidence. The workflow guides show where such a check fits in each way of working.

RunStory independently runs the software against your change and returns evidence to the coding agent. The operator cited on this page runs the factory that builds RunStory. Your team keeps the final release decision.

Join the RunStory alpha →

FAQs

Which companies run a dark software factory?

StrongDM is the one company on this list whose account matches a dark software factory. Its AI team published two rules, under which people neither write nor review the code. Other companies may run one and not have written about it, because the list covers only published accounts.

How large is a typical software factory?

A typical software factory has no standard size. The examples on this list run from one operator with a queue to companies where agents account for most pull requests. What the examples share is agents that build and a check that the builders do not control.

Do software factories work on existing codebases?

Software factories can work on existing codebases, by the accounts on this list. Stripe's engineering blog describes agents that work in the company's existing codebase with the same developer tools as its engineers. Tests that already exist give those agents a check, because a test failure with no automatic fix goes back to the agent.

What do the published factories have in common?

The published factories have two things in common. Agents build with no person at each step, and a check that the builders do not control comes before a change ships. The factories differ in who makes the last check, which is a human reviewer at some companies and an automated check at others.

Are software factories only for large companies?

Software factories are not only for large companies. One example on this list is a factory of one, in which a single operator types short requests and tries what comes out. The same two parts appear at both sizes, which are agents that build and a check that the builders do not control.