The honest near-term picture is a dark workshop, not a dark factory

AI is getting very good at coding the small problem, and we keep mistaking that for the large one.

Every impressive AI-coding demo is greenfield, and that’s no accident. A fresh app is the easy case. No pre-existing interdependence. No shared domain model owned by another team. No contracts the author doesn’t control. No legacy black box in the diagram. It’s the cheapest place to make something that looks like magic, so the demos live there.

It’s the old distinction: programming in the small versus in the large. In the small is velocity inside one bounded context: generate correct code, fast. That’s where almost the entire AI-coding conversation sits today, and it’s genuinely valuable. Prototyping, internal tools, the node-internal work of a single service. Real value, and I’m not waving it away.

But it is a bounded regime. In the large is something else: a system of many interdependent subsystems, owned by different teams, on different release cycles, at different maturities, evolving over years. The hard problem there was never generating code inside one subsystem. It’s governing change across the interdependencies between them. Faster code generation inside a node does nothing for the space between nodes.

Regulated enterprises don’t get to live inside the node. Their hardest obligations span subsystems. Two services must agree on transaction state. Data residency has to hold end to end. The audit trail crosses a dozen systems. A decision in one subsystem narrows what another is allowed to do, and its owner may not even know the choice was made. Cross-boundary change is political before it’s technical.

Be precise about what AI can and can’t do today. The limit isn’t that the models aren’t smart enough. It’s that the constraints governing a cross-boundary change live in other teams’ knowledge, ownership, and clearance authority. The model can generate subsystem B’s code. It cannot know that subsystem A’s owner changed the shared contract last sprint, and it has no standing to assert, on B’s behalf, that B is unaffected. That gap doesn’t close when the next model ships. It closes when someone decides who owns the edge.

Which is why the Dark Software Factory, the lights-out, fully-autonomous software org people keep predicting, is further off than the noise suggests. A factory is a composition. Many stations, handoffs, governance across the whole floor. That is programming in the large. A workshop is one bench, one bounded job. In the small. AI can already run the bench unattended.

So the honest near-term picture isn’t a dark factory. It’s a dark workshop. Small, real, lights-off, and no factory, because the factory floor is exactly the governance-across-owners that no model currently holds.

That floor is where the conversations I’m having with serious enterprises are landing now. Not on whether the bench can run unattended. On how you govern the edges. The model names the constraint; the platform holds the clearance, so a crossing trips a control.