Steve Yegge has a warning for teams building AI software factories: pace the agents to the people who have to live with their output. A software factory, in this context, is a collection of AI agents that build and support a product.
The bottleneck is everything around the agents
Yegge runs a factory for his online game, with roughly a dozen people involved and a few hundred active longtime players. Even at that scale, he says, the agents overwhelmed players with features, his accountant with emails, his merge queue with tickets, and his machines with load. They upset administrators, made messes in the game, and released broken open-source software without permission.
His conclusion: generating work faster does not mean the surrounding engineering organization or business can absorb it faster. He says his community’s patience gave him room to learn to slow down.
Most teams probably don’t need a factory
Factories need considerable human leadership, money, and ongoing curation, Yegge says. He abandoned five previous versions when rewriting them became cheaper than fixing them. He is not recommending swarms for everyone.
For teams that do need one, he favors building rather than buying. His factory has become tightly intertwined with his game’s operations, because it supports the product as well as building it. That is his experience with one project, not evidence that every organization should build its own system.
His prediction: everyone brings their own agents to work
Yegge expects personal chief-of-staff agents and teams of helpers to become common, with agents coordinating across colleagues and creating new workplace friction. He predicts nearly everyone will work with AI employees next year. That is a forecast, not a result from his game. His near-term advice is to learn how to work well with your own agents before scaling up.