Back to Notes

Explainer

What is an AI software factory?

A plain explanation for readers new to the idea: what a coding agent is, why one agent on a laptop is not enough, and what a factory adds around it.

Nils Widal · · 6 min read

Many people now hear the words "AI software factory" and are not sure what they mean. This note explains it from the start. It assumes you know what software is, and not much more.

First, three words

A coding agent is an AI program that changes code. You give it a task in plain words, such as "add a field for the patient's preferred language". It reads the code, writes the change, runs the tests and reports back. Claude Code is one example.

A ticket is one piece of planned work, written down in a tool such as Linear, Jira or GitHub Issues. Teams use tickets to agree on what to build and who does it.

A pull request is a proposed change to the code. A person reads it, asks for fixes if needed, and then merges it. Merging means the change becomes part of the real code.

The simple way: one agent on one laptop

Most teams start simply. Each engineer runs a coding agent on their own laptop and watches it work. This is a good way to learn. It also has limits that grow with the team.

  • Nobody sees what all the agents do, what they cost, or what waits for a person.
  • Each engineer sets the agent up differently, so results differ from person to person.
  • The agent runs with the engineer's own keys and access, on a machine that also holds email and other work.
  • The engineer has to sit with the agent. When the laptop closes, the work stops.
  • A mistake one agent makes is seen by one person, and repeats on the next laptop.

There is also a measurement problem. In a 2025 randomized trial by METR, a research group, experienced developers took 19% longer on tasks in their own projects when they could use AI tools, while they believed they had been about 20% faster. METR's follow-up in 2026 pointed toward a speedup instead, but its results were not certain. Studies disagree, and a feeling is not a measurement. Without a record of its own runs, a team cannot tell what is true for it.

What a factory adds

An AI software factory is the setup around the agents that makes their work steady, safe and visible. In our product, Widal.Factory, it has five parts.

  1. A machine per engineer, in your cloud. Each engineer gets their own virtual machine in the company's own cloud account. Their agents work there, not on a laptop. One person's agents never touch another person's work.
  2. A written workflow. The steps from ticket to merge are fixed: read the ticket, make a branch, write the change, open a pull request, answer review, merge. The agent does the parts that need judgment. The rest is plain code that does the same thing every time.
  3. Checks. Your tests and CI run on every change. CI, short for continuous integration, is the set of automatic tests a team runs on each change. An AI reviewer reads the change too. A failed check sends the work back to the agent.
  4. A person merges. Nothing reaches your main code without a person's review. The agent can propose. Only a person can accept.
  5. A hub. One page shows the whole team: what each agent is doing, what waits for a person and who has to act, and what each merged change cost.
The agent is the same one an engineer could run on a laptop. The factory is everything around it: where it runs, what steps it follows, what checks it must pass, and who decides.

How a ticket moves through it

  1. An engineer hands off a ticket in Linear or GitHub Issues, for example by adding a label.
  2. The factory starts one coding agent on that ticket, in its own branch, with your repository and your rules.
  3. If the ticket is unclear, the agent asks on the ticket and waits for the answer instead of guessing.
  4. The agent opens a pull request. Your CI runs on it, like on any other change.
  5. The engineer reviews. Comments go back to the same agent, which fixes and pushes again.
  6. The engineer merges. The factory closes the ticket, cleans up, and records the time and the cost.

What the team sees

The hub is the part people use every day. It answers three questions. What are the agents doing right now? What waits for a person, and which person? What did the work cost, and how often did a change need another round of review? Engineers open their own machine from the hub after they sign in with the company's normal sign-in.

These numbers matter. Google's DORA 2025 report, a survey of about 5,000 people in software, found that AI makes strong teams stronger and weak processes worse. A team that can see its own numbers can fix the process where AI does not help.

What it is not

  • It is not a new coding agent. It runs existing agents, Claude Code today.
  • It does not merge on its own. A person reviews and merges every change.
  • It does not move your code to us. It runs in your own cloud account, under your own agreements with the model provider.

What runs today

Widal.Factory runs today on AWS, with Linear or GitHub Issues as the tracker and GitHub for code. Jira is next. Google Cloud and Azure are not built yet. We are testing a way to describe any workflow as a file, with checks on each step, so a team can add its own workflows without code changes.

The same base, with workflows, checks and a person's sign-off, can later carry work outside engineering. That is planned, one department at a time.

Read next: how it works, step by step, and Widal.Factory for engineering teams.

Sources

  1. Anthropic, Building effective agents
  2. METR, randomized trial of AI tools and experienced developers (arXiv 2507.09089)
  3. METR, uplift update (February 2026)
  4. DORA, 2025 State of AI-assisted Software Development

Tell us which work you want to hand off first.

Book a call