Dark Factories 🏭

Happy Friday,

Talk to 10 people about how they use AI and you get 13 or so answers back. Lately, Martin Fowler's site has published several interesting articles about things like the cost of refactoring and whether TDD is useful with agents.

I'll summarize them as: the cost for agents to work in well-factored code is significantly cheaper. TDD adds token cost, but doesn't appear to improve the code itself. The last one needs a small clarification. TDD does several things to codebases, and one of them is strongly encouraging different design choices. Agents don't respond to that.

Why does this bit about quality matter? Well, there are folks experimenting with factories, dark factories, and superdark factories. Some of you might not be aware of these, and some of you are experts.

All of these terms represent autonomous work done by agents, but each adds a unique level of freedom and problem-solving. I'll give a really simple way to understand the differences:

• Factories: Accomplish this task for me and tell me when you're done.

• Dark Factories: Accomplish this goal within these boundaries; figure out your own tasks and do them.

• Superdark Factories: Set your own goals and boundaries, then accomplish them.

Each type of factory shares the characteristic that a human is no longer really steering what is going on. In the factory, the human waits for the factory to produce its output. In the dark factory, the human simply hopes to see the goal accomplished, not the detailed execution. Superdark factories are cutting-edge, but this would be more like a friend telling you what they've been up to.

Some folks are running toward these concepts; others are running away from them.

I want to offer a thought experiment. What would it take for you to operate a dark factory? I bet some of you are grappling with a loss of control, the risks of mistakes, lack of visibility, etc. This is the thought experiment. Try to solve these things.

Now, some of those answers will be very specific to building your factory, like what specific MCP servers you'll need. Most, however, will be answers that would have always worked in your teams.

The factory idea is also about building high-performing teams. Sure, agents and people are vastly different, but the leadership isn't. Neither is the refinement of tooling and processes to create a stable and safe work environment.

I'll close out with a nugget from a recent conversation with an engineering manager experimenting with dark factories: "There is simply no excuse for software to have problems at this point."

Sincerely,

Ryan

PS: Factories are a fun idea, but with them come the risks of producing stuff nobody wants in amounts nobody can handle. Watch my talk on overproduction and how much it costs you.

PPS: Enjoy this newsletter? You should share it with someone. Also, if you want to talk about factories, you can just reply or schedule a time.