01 · The Game That Was Training
A generation drilled for a job that did not exist yet. #
There is already a thing called the Factorio effect. Ask anyone who has lost a weekend to the game and they will tell you it is the compulsion, the just-one-more-belt loop that turns midnight into dawn, the way the factory crawls into your head and starts asking to grow while you are trying to sleep. That is the famous version. It is not the one I mean.
The version I mean is what the game was quietly doing to the people who played it. Factorio is a few hundred hours of training in a single discipline: how to run a system of autonomous producers that you do not personally operate. You do not place the iron on the belt by hand. You build the machine that does, and the machine that feeds that machine, and then you step back and supervise a thing that runs without you. You learn to think in throughput instead of tasks, in bottlenecks instead of effort, in blueprints instead of buildings. You learn that the factory is the unit of work, not the wrench.
That is the exact skill the AI software factory now demands, and almost nobody was taught it on purpose. The people picking up agent orchestration fastest are not the ones with the most LLM theory. They are the ones who already spent a year of evenings learning, in their hands, that you do not do the work, you build the thing that does the work, and then you keep the thing alive. The game was a flight simulator for a cockpit that had not been built yet. Now it has been built, and the simulator hours transfer.
The famous Factorio effect is the addiction. The real one is the training. A generation learned to supervise autonomous production for fun, years before the job existed.
02 · Played, Not Designed
You do not architect a factory. You play through one. #
Here is the thing the game teaches that no amount of upfront thinking can replace. You cannot design the right factory from the armchair. You can sketch a plan, and the plan will be wrong, because the problems that matter do not show up on paper. They show up when the belts are running and something backs up two hundred tiles upstream for a reason you could not have predicted until you watched it happen.
So you build a first base. It is spaghetti. Everyone's first base is spaghetti, by law, not by failure. You learn what it teaches you, and then you do the thing that separates the people who get good from the people who stay stuck: you scrap it and start from zero, and the second base is cleaner because the first one taught you where the pain lives. Then you scrap that one too. I am on iteration seven of the real factory. Seven from-scratch rebuilds, each one carrying forward only the lessons, never the structure. That is not a confession. That is the mechanic.
Plan to throw one away · Brooks, The Mythical Man-Month, 1975
"Plan to throw one away; you will, anyhow." The first system you build is the one you learn on. The second is the one that works, because the first showed you the shape of the problem you could not see before you built it. Fifty years before iteration seven, Brooks named the loop and told you to expect it. Factorio just turns it into a save file. The first base is supposed to be torn down. Keeping it out of sentiment is the actual failure, not the teardown.
This is why ground before you theorize is law and not a slogan. In Factorio there is a tech layer you genuinely cannot understand until you have played up to it. Read every wiki page on the late-game and it will still not be real to you, because the constraints of that tier only bite when your factory is large enough to feel them. The real factory is identical. The problems of running a hundred concurrent agents are not visible from the design of running ten. You play through to them, or you do not know them at all.
The first base is spaghetti by law. The right factory is not the one you designed. It is the seventh one, found by scrapping the first six.
03 · The Agent Is on the Belt
The agent is the iron plate, not the assembler. #
This is the inversion that makes the metaphor more than a cute parallel, and it is where most people have the picture exactly backwards. The intuition is that the agent is the machine: the assembler standing on the factory floor, the thing that does the work. It is not. In a factory built right, the agent is the thing on the belt. It is the iron plate. It rides in, gets consumed at its station, and is dumped.
The durable machinery is something else entirely. It is the substrate: the dispatch primitive, the chain runner, the supervisors, the belts themselves. That layer is fixed, owned, and long-lived. The agents are not. An agent is spawned for one transformation, it does the one thing, it emits, and it dies. You do not keep it. You do not grow attached to it. You do not ask it to remember. It was a plate, and plates do not persist past the machine that consumes them.
And this is the part the industry is killing itself over: that design is exactly what dissolves the context problem. An iron plate does not remember being ore. It carries no history of the furnace. The state does not live in the item being consumed, it lives in the artifact that rides the belt forward: the brief becomes code, the code becomes a review, the review becomes a merged PR, each one handed to a fresh consumable that transforms it and is thrown away. Nothing accumulates inside a worker because no worker lives long enough to accumulate. There is no drift, because there is nothing for drift to happen to. The whole industry is trying to give the plate a memory: longer context, vector stores, summarization, a scratchpad the worker reads back to itself. The factory builder does the opposite. Keep the plate moving. Let the belt carry the state.
Everyone is trying to give the iron plate a memory. The factory builder keeps the plate moving and lets the belt carry the state.
04 · Two Streams on the Floor
Agents are consumed. Artifacts compound. #
Once you see the agent as an item, you can see that there are actually two kinds of thing moving across the floor, and they have opposite life spans. The agents are the disposable stream: consumed, dumped, gone, the cheap labor that exists for one transformation and never again. The artifacts are the other stream: the code, the briefs, the reviews, the merged work. Those persist. Those ride forward. Those are the point.
And the artifacts do something the agents never could. They compound. In Factorio you do not mine iron to make iron. You mine iron to make the machines, and the machines make better machines, and eventually the factory is producing the components to build a second factory, and the second one is faster because the first one taught you the blueprint. The output loops back into capacity. That is the endgame of the game and it is the endgame of the real thing: a factory whose product is more and better factory.
This is why the continuous improvement is not a phase you finish. The artifacts that compound include the factory's own machinery: the doctrine, the supervisors, the blueprints. The system refactors itself with its own output. The agents that built this week's improvement are already dead. The improvement remains, on the belt, feeding the thing that builds next week's. Disposable labor, compounding capital. The workers are cattle and the artifacts are the herd that grows.
The agents are consumed and forgotten. The artifacts ride forward and compound. Disposable labor, compounding capital, and the factory's best product is more factory.
05 · The Map Zoom
You place one belt and end up over a continent. #
Watch anyone play Factorio from start to finish and the most honest record of their progress is the zoom level. Hour one, the camera is down on the floor, and you are placing a single belt segment and a single inserter by hand, tile by tile. Hour two hundred, the camera is so far up that individual machines are invisible and you are moving entire districts at a time, reasoning about a city the size of a small country, and you have not personally placed a belt in days.
Your altitude rises as the machine fills in beneath you. That is not a metaphor I am imposing on the game. It is the literal camera, and it tracks the literal change in what you are doing. You start as the worker placing the part. You end as the operator moving the blueprint. The work did not disappear. It moved below you, into the automation, and you moved up to where the leverage is.
I have a name for this from before I had the Factorio frame for it: humans moving upchain. The agents move in below, into the implementation slot, and the human relocates above, to intent and structure and judgment. Factorio renders the whole shift as a camera move you can feel in your hands. The moment you stop placing belts and start stamping blueprints is the moment you moved upchain, and the game makes you feel exactly when it happens, because the zoom level changes and you never go back down.
The zoom level is the most honest record of progress. You start placing belts by hand. You end moving districts. The altitude rises as the machine fills in below.
06 · The Blueprint Fractal
Once a unit works, you stamp it. #
The first time you solve smelting in Factorio, you sweat it. You agonize over the ratios, the belt lanes, the inserter placement. And then it works, and you select the whole thing, and it becomes a blueprint, and from that moment forward you never solve smelting again. You stamp it. Need ten times the throughput? Stamp it ten times. The hard-won unit becomes a tile you place without thinking, and your attention moves up to the next unsolved thing.
This is the same shape as the reusable core I described in The Capability Bus: solve the thing once, correctly, then turn it into something you instantiate instead of rebuild. The blueprint is the compiled unit. What changes between stampings is small and declarative, the inputs and the placement, while the internals stay fixed. You do not write twenty smelting arrays. You write one and stamp it twenty times, which is exactly the move from twenty hand-built integrations to one core and twenty manifests.
The self-similarity is the tell that you have built substrate and not a one-off. A blueprint that only works in the spot you first built it is not a blueprint, it is a building. A real one stamps anywhere, composes with copies of itself, and scales by repetition rather than by redesign. When your factory starts to look the same at every zoom level, the same pattern of pattern of pattern, that is not an accident of style. It is the proof that the unit was actually general, and generality is the thing you were trying to build the whole time.
A blueprint that only works where you first built it is not a blueprint. It is a building. The real one stamps anywhere and scales by repetition, not redesign.
07 · Find the Bottleneck or Optimize Nothing
The whole game is one question, asked forever. #
Strip Factorio down to its core and it is a single question repeated until you stop playing: where is the constraint right now, and what is it. Every productive hour in the game is spent finding the one belt that is starved, the one machine that is the wall, the single place where the whole line is waiting. You fix that, and the constraint moves somewhere else, and you go find it again. Optimize anything that is not the current bottleneck and you have done precise, satisfying, completely worthless work. The line is exactly as fast as its slowest point and not one item faster.
The Theory of Constraints · Goldratt, The Goal, 1984
A system's throughput is set by its single binding constraint, and improving anything other than that constraint is motion without progress. Goldratt taught it as a business novel about saving a manufacturing plant by finding the one bottleneck and subordinating everything else to it. Factorio is that book rendered as a toy you cannot put down. The discipline ports straight to the agent factory: your throughput is gated by one thing right now, the credential pool or the review lane or the merge gate, and the work that matters is finding which, not optimizing the six that are already fast.
And the stakes scale with the setting. That base is death world marathon: every recipe costs several times more, and the map is trying to kill you the entire time. On that setting a bottleneck is not an inefficiency, it is a cause of death. If ammo production is the constraint and you miss it, you do not lose throughput, you lose the wall, and then the base. It is why half the districts in the shot are military, a whole war economy whose only job is to buy time for the production economy to outgrow the threat. Goldratt in a boardroom is about quarterly numbers. Goldratt on death world is about whether you survive the next wave, and the real factory is closer to that than anyone likes to admit: a bottleneck in the part of the pipeline that keeps you safe is not a slowdown, it is an incident.
And here is why the bottleneck obeys runtime proof beats repo text. You cannot find the constraint by reading the blueprint. The design looks balanced on paper, every ratio correct, and then you run it and the items pile up in a place the math said was fine, because reality had a detail the design did not. In the game you watch where the belt backs up. In the factory you read where the work backs up. Either way the constraint is discovered by running the thing, never by reasoning about it, and anyone optimizing from the diagram instead of the live floor is polishing a machine that was never the wall.
The line is exactly as fast as its slowest point. You find that point by running the factory, never by reading its blueprint.
08 · The Factory Must Grow
It is a practice, not a project. #
The community has a phrase, half a joke and half a creed: the factory must grow. It is the thing you mutter when you should be asleep and instead you are extending the rail network one more time. And underneath the joke is the truest thing the game knows about the work, which is that the factory is never done. There is no win screen where you set down the wrench and the thing is finished. There is only the next constraint, the next blueprint, the next rebuild, the next zoom level out.
That is the honest shape of the real one too. Iteration seven is not the last iteration. There will be an eight, and it will tear down structure that took weeks to build, and that is correct, because the structure was never the asset. The lessons were the asset, and the artifacts that compound were the asset, and the practice of building, scrapping, and rebuilding at rising altitude was the asset. The factory is not a thing you complete. It is a thing you operate, and operating it well means it is always, by design, in the middle of becoming a better one.
So the two Factorio effects turn out to be the same effect seen from two distances. Up close it is the compulsion, the factory crawling into your head and asking to grow. Stepped back it is the training, the discipline of running autonomous production until it becomes the way you think. The game taught a generation to feel both, for fun, on nights and weekends, years before there was a factory worth the name to build. Now there is one, and the only surprising thing is how few people noticed they had already been to the school.
There is no win screen. The factory is not a project you finish. It is a practice you operate, always in the middle of becoming a better one.
Factorio was right about how to build a factory before the industry needed to know.
It encoded the whole thing in mechanics, and the mechanics turned out to be right. A factory is played, not designed: you build the wrong one six times to find the seventh. The agent is the item on the belt, not the machine: consumed, dumped, carrying no state, which is the move that dissolves the context problem the rest of the field is still paying for. The artifacts compound while the workers are forgotten. The altitude rises as the automation fills in below. The unit that works gets stamped, not rebuilt. The constraint is found by running the line, never by reading it. And it is never finished, because the factory must grow.
None of that was on the box. It was just how the game worked, and how the game worked turned out to be how the real thing works, because a factory is a factory whether the items on the belt are iron plates or agents. The people who will build the real ones best are the ones who already have the hours in, and most of them do not yet know that the weekend they lost was the apprenticeship.