Software Factories Run on Pre-Prod

Every impressive self-driving software factory demo has something in common: the software has no customers yet.

Autonomous agent loops are real and they are getting better fast. Chain a builder agent to a critic agent, have a third agent adjudicate, run it until everything resolves. I find this architecture genuinely exciting, and we are building toward it. But watch where these loops actually run today. They run on greenfield projects, internal tools, and pre-production code. Places where the blast radius of a bad merge is zero.

Now look at a typical engineering org. A thousand tickets in the backlog, a production system, paying customers on the other end of every deploy. You cannot point an autonomous loop at that backlog and let it rip. Not because the agents are not capable enough, but because the risk math does not work. One bad autonomous change in production costs more than a hundred good ones save. Teams with real customers know this in their bones, which is why the fully autonomous factory remains a spectator sport for most of them. It is fun to watch. It is not practical for their reality.

The pattern that does work in production is narrower and less cinematic. Context in, background execution, reviewable output, human approval. The loop can be as aggressive as you like right up until the merge decision, and then a human owns the yes or no. That gate is not a concession to weak models. It is what makes aggressive automation deployable at all, which is the argument I made in human review is not a limitation. The gap between the loop demo and the loop in production is the same gap that has always separated the demo from the deployment.

The factories will earn their way into production eventually, one workflow at a time, as trust accumulates. Until then, the practical question for a leadership team is not whether to run autonomous loops. It is where the approval gate sits, and who holds it.