The factory is not yet reproducible

The millionth part rolls off a production line without much ceremony. By then, thousands of engineering decisions have become invisible. They're buried in fixtures, machine settings, control logic and the judgment of people who know the line.

Now build the same line in another factory. You can buy the equipment again. You can copy the drawings. But you cannot simply copy all the work that made the first line reliable.

The product is reproducible. The factory largely isn't.

Manufacturers routinely reuse equipment designs, layouts and commissioning procedures. What remains difficult to transfer is the knowledge accumulated while getting everything to work together. Some of it is documented. Some lives in custom fixtures, exception-handling code and adjustments that took months to get right.

Consider something as ordinary as moving a component between two machines. The movement itself takes seconds. Automating it may require a custom fixture, precise positioning, safety controls and a way to handle the part arriving slightly out of place. Multiply this across hundreds of operations and a production line becomes a collection of decisions that don't travel easily.

I saw a version of this problem while advising Multiply Labs, which builds robotic systems for manufacturing advanced biological therapies. Much of the underlying science already happens inside specialized instruments. The difficult work is also between those instruments: transferring material, preparing consumables, starting the next operation and recording what happened. Sterility, timing and traceability matter at every handoff.

Multiply Labs uses modular robotics to operate existing instruments rather than requiring manufacturers to redesign every scientific process. But modular doesn't mean universal. Robots still need to interface with specific equipment and handle materials in ways the equipment can accept. Standardizing the automation doesn't eliminate the physical particulars; it changes how often you have to solve them.

Traditional industrial automation handles those particulars by removing variation wherever possible. Fix the part in place, control the environment and program the movements. This works extraordinarily well, but a new component, instrument or layout can trigger another integration effort.

Much of automation engineering goes into handling exceptions. A part arrives misaligned, a gripper misses its pickup, or a component doesn't seat correctly. Engineers design fixtures, tune sensors and program recovery logic for these cases. Physical AI could reduce how much of that work must be repeated, using learned perception and control to handle variation across different installations. The advance isn't a robot performing a new task, but making that capability easier to deploy elsewhere.

The point isn't simply to make machines easier to rearrange. It's to carry forward the engineering that made them useful. A capability developed for one production line should require adaptation, not reinvention, when deployed elsewhere. Otherwise, the hardware may be modular, but the engineering remains largely bespoke.

But reusing a movement isn't the same as reusing a process. The system still needs to coordinate equipment, verify outcomes, handle failures and satisfy safety and quality requirements. Modular hardware, common interfaces and conventional control software matter as much as the intelligence inside the robot.

The real test is what happens at the second installation.

If the first deployment takes months of engineering, how much must be repeated? Can the next system retain what was learned about the process and adapt only to what changed? And what about the tenth or twentieth installation? If the engineering effort remains roughly the same each time, we've made better machines without fundamentally changing how manufacturing scales.

That changes the math for smaller production runs. Custom automation often needs high production volumes to justify its upfront cost. Reusable capabilities could make smaller production runs and specialized processes viable. Over time, easier deployment might also make it practical to establish production closer to where products are needed, rather than concentrating every process in facilities large enough to absorb the engineering expense.

Equipment wears, materials vary, sites differ and regulated processes still need validation. Some jobs will always be better served by purpose-built machinery. The goal isn't a factory that can make anything.

We know how to reproduce the product. The hard problem worth tackling is reproducing everything that makes the product possible.