Generation is the visible ten percent
AI can produce a plausible patch in minutes. That is useful, but a patch is only the first artifact in a much longer chain: understand the outcome, change the right system, prove the change, deliver it, and verify that the real target state moved.
Most agent demos stop at the moment code appears. Real work starts immediately after that moment. Someone still has to preserve intent while tests fail, reviews find edge cases, deployment changes the environment, and a retry threatens to forget why the work existed in the first place.
Factory Next is an operations experiment
Factory Next is in active development. The working question is simple: can a small set of autonomous coding agents behave like a dependable delivery operation instead of a collection of clever chat sessions?
The system treats ownership and evidence as product behavior. Work has one accountable lane. Milestones stay attached to the original outcome. Independent review examines the exact candidate. A deployment is followed by an acceptance check rather than a victory message.
- One owner for each outcome, so responsibility does not dissolve between agents.
- Evidence attached to the exact code that was reviewed and delivered.
- Retries that return to the original user intent after a technical failure.
- Acceptance based on a measured target state, not activity inside the pipeline.
The economic reason to care
Cheap code generation does not automatically create cheap software delivery. Rework, supervision, escaped defects, and abandoned half-finished tasks can consume the savings. The useful unit of automation is not tokens or patches. It is an accepted outcome per unit of owner attention.
That is the metric I am building toward: less time spent babysitting the mechanics, without pretending that uncertain work is finished. I will share the design choices, failures, and measurements here as the system develops.