Skip to main content
FeatureFactory

The FeatureFactory manifesto

The factory was never the insult. Shipping without instruments was.

Build repeatably. Use agents aggressively. Keep engineers in control. Own the output. And measure what happens after the work ships.

Read the argument

“Feature factory” is supposed to be an insult.

It describes a team that cranks out features nobody asked for and never checks whether they worked. Output with no outcome. Motion with no instrument panel.

But the problem was never the factory. A good factory is repeatable, observable, and relentlessly improved. It preserves what works, exposes where the line fails, and makes quality a property of the system instead of a heroic final inspection.

The problem is a factory with no instruments-one that ships and ships, counts what left the line, and refuses to look at what happened next.

AI makes that distinction urgent. A new generation of software factories can produce more code at extraordinary speed. Some sell that speed to executives, move judgment behind a service wall, and return a number the engineers responsible for the system cannot verify.

More code is not the goal. More useful, durable change-with less avoidable effort-is.

What we believe

01

Output is not an outcome

Lines written, tickets closed, and pull requests merged are activities. The work earns its keep only when quality holds and the customer or business condition changes.

02

AI raises the burden of proof

When producing plausible code gets cheaper, counting code gets less meaningful. The faster the factory runs, the more carefully we must inspect durability, rework, and the result.

03

Engineers keep the judgment

Agents can search, draft, transform, execute, and test. People still own product choice, architecture, risk, review, and the definition of success.

04

Ownership is a product feature

Code, plans, prompts, configuration, and measurement history should increase a team’s capability-not become leverage that keeps it inside a vendor.

So here is the bet: build like a factory-planned, agent-driven, repeatable. But measure like a scientist.

Set the baseline before changing the process. Score the work and show the rationale. Separate known authorship from estimates. Tie the plan to the change and the change to the result. When the metric did not move, say so. When the speed created rework, count it.

A system like that does not diminish engineers. It gives them better instruments for deciding where judgment matters and where automation can safely compound their work.

A factory becomes dangerous when it hides the line. It becomes powerful when the people operating it can inspect every station.

That is the version of “feature factory” worth reclaiming: not an organization optimized for throughput at any cost, but a system that makes intent, execution, quality, ownership, and outcome visible in one loop.

Our product contract

Three commitments we refuse to compromise.

Developer-led

Adoption must make sense to the people responsible for the code, not only to the executive approving the contract.

Transparent

Every important score needs a rationale, source, definition, and limit a skeptical engineer can inspect.

Customer-owned

The team keeps its code, data, artifacts, and exit path. The platform should be useful without becoming captivity.

What we refuse

The line matters as much as the output.

  • We will not call more generated code a productivity result.
  • We will not present estimated AI attribution as ground truth.
  • We will not hide product judgment inside an “autonomous” workflow.
  • We will not make customer IP the price of using the factory.
  • We will not celebrate a shipment before reading the outcome.

Build the version worth reclaiming

Put your next feature on an outcome loop.

Join early access to plan from evidence, build under engineering control, and measure what the change actually did.

Read-only to startNo replacement rolloutYour IP stays yours

We’ll use this only to coordinate early access.

Your baseline comes from your data-not a generic benchmark.