SCADA was the baseline for the factory that could not think. BASE is the baseline for the factory that can.

Walk into a factory today and you will find, somewhere near the back, a room full of screens. On those screens are the dials, gauges, and alarm lists of everything happening on the floor. A person sits there and watches. When something drifts, they act. This room has a name — SCADA, Supervisory Control and Data Acquisition — and for thirty years it has been the unspoken baseline of industry: the plant is dumb, the intelligence lives in one room, and a human supervises the gap between them.

It was the right design for its time. Sensing was expensive, compute was scarce, and the only place to concentrate intelligence was a control room. Everything since — the MES, the historian, the dashboard — has been a better version of that same room.

The factory that is arriving does not have that room at its centre, because it does not have a centre. Picture it instead as a society. Every machine has a small program that is that machine in software — it knows what it is, what it can do, what it is doing right now, and what it did last shift. Every order that comes in the door is a program too, one that wants something built by a date, for a cost. Every cell, every crew, every workshop is a program. None of them waits for the room. The order program asks the floor, in effect, who can cut this tonight, and what will it cost me? — and the machine programs answer, bid, and commit. Work is not dispatched from above. It is negotiated among peers, continuously, while the plant runs.

This is not a dashboard over the old plant. It is a different kind of plant: one made of thousands of small, living, autonomous programs that sense, decide, and act on their own — and that keep thinking a kilometre underground when the network drops, because the intelligence lives at the edge, in each program, not in a room that has stopped watching.

Science fiction reached this picture before industry did. Charles Stross's Accelerando imagined an economy run by software agents transacting faster than any human could follow. Mass Effect's Geth were not a single hive mind but many programs reaching consensus — no CEO, no central brain, and it still cohered. The fiction was pointing at something real, and it was right about the shape of it. It was wrong about the ending, and we will be precise about that later, because the difference between the mechanism and the myth is exactly where most "AI for factories" pitches quietly lie.

The shift is as large as the one SCADA made when it replaced men walking the line with pushbuttons. SCADA answered how do we watch a plant? The next baseline answers a harder question: how does a plant run itself? That baseline is what this paper is about. We called it BASE.

SCADA answered how do we watch a plant? The next baseline answers a harder question: how does a plant run itself?

There is a belief among engineers, AI labs, and today's automation stacks that intelligence is a question of scale: more compute, more context, more power, and a complex system will emerge. A bigger, smarter central brain that finally sees everything and executes it all correctly. Herbert Simon spent a career explaining why that brain cannot exist. His central idea, bounded rationality, won him the Nobel Prize in 1978, and it is deceptively simple: every decider, a person, a control room, a program, is boxed in by what it can know, how fast it can compute, and how much attention it can spare. Perfect optimisation assumes an actor with unlimited information and unlimited time to use it. No such actor stands on a factory floor, in a mine, or anywhere in industry. The central brain does not fail because we have not built it well enough. It fails because the task we hand it, see everything, weigh everything, decide everything, fast enough to matter, is not a solvable problem. It is a category error.

If no one at the centre can hold the whole picture, the decisions have to move to where the knowledge already lives: to the machine that knows it is running hot, the order that knows its own margin and due date, the cell that knows what remnant is left on the rack. Friedrich Hayek reached the same conclusion from the other side in 1945: the knowledge in an economy is irreducibly local and dispersed, and the job of a coordination system is not to gather it all in one place, which is impossible, but to let local actors signal and adjust. Simon arrived through the limits of the mind, Hayek through the limits of the market. We ask more of our industrial base every year: mass production and custom production at once, under conditions that change while the work runs. No closed-form solver answers that. To insist on one is to fight physics. Nature never tried. The systems that scale without a central mind, a nest, a body, a city, an ecosystem, solved it the same way, through emergence: simple parts acting locally, composing into a coherent whole. That is the model industry has been missing, and it has a precise name.

So the future plant does not control every resource from a central point. It declares what outcome it wants and negotiates it according. An order announces what it needs; the machines that could do it bid; a contract is struck and the work is awarded — all at runtime, with no central scheduler holding the pen. This is not a metaphor borrowed from economics; it is a named, forty-year-old engineering protocol. Reid Smith described the Contract Net Protocol in 1980: a manager announces a task, contractors bid, a contract forms, no single controller required. It is how distributed systems coordinate without a dispatcher, and it is the model BASE is built toward. The hard part — the wiring that lets one agent speak directly to another, peer to peer, with no server in the middle — is already in place. The contract-net negotiation that rides on top of that wiring is what I am building now.

Science fiction drew the shape before industry built it. The Geth of Mass Effect are the cleanest popular picture: not a single hive mind, but many small, specialised, deterministic programs that pool processing and reach consensus, each contributing one component that pushes the outcome toward the goal. No fused consciousness, no one giant AI. What the fiction gets right is the distributed architecture. What it misses is the hard part: dynamic hierarchical depth. There is hierarchy, and there has to be a north-star leader, but never only one. There is one for every goal, order, and task the system takes on, and the resources beneath them are shared, so they must be negotiated for. There is an older word for this, and a better one: a holon, a thing that is at once a whole and a part. Arthur Koestler coined it in 1967, and it is the root of holonic manufacturing systems, the field I spent a doctorate inside. A machine agent is a whole, because it governs itself, and a part, because it serves the order. It never dissolves into a hive, and it never becomes the dictator.

This is the point where a careful reader gets nervous, because the story is starting to sound like a pure market, every machine a merchant, every order a buyer, the plant a trading floor. Simon warned against exactly that overcorrection. In "Organizations and Markets" (1991) he pointed out the fact that embarrasses market purists: most economic activity does not happen across markets at all. It happens inside organisations, inside hierarchies of coordinated decisions. The future plant is not a Swiss watch, a rigid mechanism of precisely-fitted parts, which is what SCADA tried to be, and it is not a spot market of atomised traders either. It is a holarchy: a nested hierarchy of near-decomposable units that coordinate mostly within themselves and negotiate only at their edges. An insect hive, not a Swiss watch — but a hive of holons, not a mob of merchants.

The future plant is not a Swiss watch, and it is not a spot market. It is a holarchy: a fractal of holons, not a single super-intelligence, not a mob of merchants.

Grant that the plant negotiates. Negotiates toward what? Here the honest answer breaks with a century of industrial ambition. The dream of the MES era was the global optimum, pour the whole week's orders, machines, and constraints into one model and solve for the single best schedule. Simon put the axe to that dream too, with an idea he named satisficing, a splice of "satisfy" and "suffice." Real deciders, he showed, do not optimise; they set an aspiration level and search only until they find something good enough to meet it, and then they act. Not from laziness, but because optimisation over a live, changing world costs far more than it returns: by the time you have computed the perfect schedule, three orders have changed and a saw has thrown a blade. A plant that satisfices keeps its promises. A plant that insists on optimising spends its life recomputing an answer that is already wrong.

A plant that satisfices keeps its promises. A plant that insists on optimising spends its life recomputing an answer that is already wrong.

There is a physical picture of satisficing, and BASE runs on it. Cool a metal slowly and its atoms settle toward a low-energy state: not the one perfect crystal, which would take an eternity to reach, but an excellent, stable, good-enough arrangement that holds. A plant is the same, and the only hard question is good enough by what measure. Every objective a plant carries is a cost. A machine that overruns costs energy to recover. One component takes more energy to cut than another. One ore takes more to move or refine. Even a delay is overhead, and overhead is energy in the end. So BASE takes the axiom that collapses the mess: every cost normalises to energy, and each holon prices any action by its ultimate energy cost.

That makes energy the medium of negotiation. A holon whose goal matters more carries a larger energy budget, and outbids a lesser one. A machine that reaches the goal for less energy wins the contract. This is arithmetic, not judgement, which is why it was rarely done: by hand, a single plan meant hours of costing and redrawn holarchies. A computer runs thousands at once. Reduce every complexity to an energy budget and the plant settles continuously toward the lowest-energy way to keep its commitments, re-settling the instant reality shifts. It never freezes on a false optimum. It anneals, in the exact sense Kirkpatrick gave the word in 1983: accept the occasional worse move so you never get trapped, and cool toward a low-energy state. Auctions of this kind, market-based control (Wellman 1993, Clearwater 1996), have run inside building and computing systems for decades. BASE points them at the plant.

Energy is not the only scarce currency. Simon named the other before almost anyone: "A wealth of information," he wrote in 1971, "creates a poverty of attention." A plant drowns in data, every sensor, every tag, every alarm, and the scarce thing is not the data but the attention to act on the right of it at the right moment. This is why BASE's cortex is, by name, an attention allocator. The plant's economy runs on the two things it cannot manufacture more of: energy and attention.

"A wealth of information creates a poverty of attention."

This is the moment to be honest about the fiction that got here first. Charles Stross's Accelerando imagined "Economics 2.0" — agents transacting over scarce resources so fast that human narrative, and eventually humans themselves, fall behind and become spectators, or worse. The mechanism is exactly right: dense populations of agents negotiating capacity through price-like signals genuinely do outperform a central planner when the world is noisy and knowledge is local. That is the part we build. The ending is exactly wrong, and we say so plainly. Markets fail — they go thin, they get gamed, a single bottleneck cell becomes a monopoly — and a factory is not a novel about the heat-death of human agency. So the economy of energy is wrapped in the things Stross's runaway markets lacked: circuit-breakers, audit trails, and a hard human acknowledgement before anything that can move metal or hurt a person. We take the mechanism. We leave the apocalypse on the shelf.

Goal-integrity is not a prompt you hope holds. It is a property of the architecture: the assurance an operator needs before a program is allowed anywhere near a machine.

Everything so far describes a future that is, on paper, buildable — bounded-rational agents negotiating an energy economy and satisficing toward kept promises. So why does it not exist? Simon saw the obstacle in 1962, in an essay called "The Architecture of Complexity," through a parable of two watchmakers. Hora and Tempus both make fine watches of a thousand parts. Tempus builds each watch as one long assembly, and if he is interrupted — the phone rings constantly with new orders — the half-built watch falls to pieces and he must start again. Hora builds his from stable sub-assemblies of ten parts, each of which holds together on its own; when the phone rings, he sets down a stable piece and loses only the part in his hand. Hora prospers; Tempus goes broke. Simon's lesson is that complex things which survive are built from stable intermediate forms — nearly-decomposable parts that each hold together while the whole comes together, so that an interruption costs a piece and not the world.

Today's agent infrastructure is Tempus. The dominant way to build software agents, a Python framework, an external scheduler bolted on the side, a vector database for memory, a cloud sandbox, was designed for programs that start, do one thing, and stop. Ask it to run forever and it hits a wall. State lives in memory and dies on a crash; a human restarts the agent by hand and hopes the situation can be reconstructed. The "phone call", an unhandled exception, a dropped connection, a context window that fills and forces a lossy hand-over, destroys work in progress, because the work was one long monolithic assembly with nothing stable underneath it. This is the Runtime Wall, and it is not a library problem that a better framework patches next year. It is Tempus's problem, and it is structural.

• SCADA/MES is the wrong baseline, too. Beyond the runtime, the reigning architecture is the wrong shape: SCADA and the MES were built to watch a dumb plant from a central room, and a central-supervisory design can no more host a society of negotiating agents than a mainframe can host the internet.
• So the good operators rebuild it by hand. The most capable new manufacturers do the only thing they can — they build their own plant software from scratch, in bespoke TypeScript, each one reinventing the same runtime badly and alone. A "FactOS" rebuilt in every greenfield shop because no forkable foundation exists. The demand could not be louder; the infrastructure simply is not there.

And the window is open now — which is why I stopped everything else. The tooling is moving so fast that the baseline for the agentic factory will be set in the next few years, by someone — the way SCADA's baseline was set once and then held for thirty. If it is set by a thousand incompatible, hand-built FactOS stacks, we inherit the exact trap we are in today, one layer up: locked-in, unforkable, obsolete the moment the vendor moves on. The standard has to be set now, deliberately, by something built from the start to be forked and owned. That is not a distant ambition. It is the reason I walked away from a decade-long mining contract to build this while it still matters.

BASE is Hora's workshop. It is built, from the metal up, out of stable sub-assemblies that survive the phone call — and that single commitment is what turns the buildable-on-paper future into a thing that runs a kilometre underground today.

Simon's near-decomposability is not only a survival trick; it is the blueprint for how BASE scales. A nearly-decomposable system is one whose parts interact strongly within themselves and only weakly across — so each part can be built, understood, and repaired on its own, yet still composes into a coherent whole. That is precisely the holon, and precisely BASE. One agent is a near-decomposable whole, with its own four sectors, its own cortex, its own biography. A cluster of agents negotiates only at its edges. An activity agent owns a fractal of sub-agents, each a complete BASE in its own right. A holarchy spans many sites under a single directive. The same primitive, repeated at every scale, each level nearly-decomposable from the ones above and below — so a fault stays local, an upgrade stays local, and understanding stays local. Get the primitive right, Simon says, and every order of structure above it inherits its guarantees. That is the entire design philosophy of BASE, compressed into one 1962 sentence.

Look closer at what each holon holds. The order holon knows how to run an order. It does not know how to paint, or cut, or bend, and it never needs to. It holds the context, the tools, and the knowledge for one job: taking the order apart and holding the whole together. At the paint step it hands the paint shop a spec and nothing more. The paint shop is the expert. It breaks that spec into its own sub-tasks and hands each to a narrower expert still, powder coating to one machine, resin spraying to another, each holding only the knowledge its own work demands. Expertise lives where the work lives, and goes no deeper than it has to.

This is the part that matters for anything built on a model. A monolithic agent tries to hold every domain at once, the order, the paint chemistry, the machine codes, in one context window that grows until its judgement decays. A holarchy refuses that by construction. Every level reasons over a small, bounded surface: its own domain, its own tools, the knowledge a specialist encoded into its behaviours once. No level carries the context of the levels beneath it, because it never needs it. This is holonics as the physics of context optimisation, built into the architecture rather than asked for in a prompt. And because each level is a bounded domain expert, its outcomes are deterministic and accountable: you always know which holon owns a decision, what it knew, and why it acted.

• The physics argument (BEAM) — Hora, made real. Erlang/BEAM was built in 1986 for telephone switches that could not go down: per-process isolation, hot code loading, preemptive concurrency across millions of lightweight processes, native distribution. Those are Simon's stable sub-assemblies enforced by the runtime - a crashed behaviour cannot corrupt its neighbours, and an interruption costs the part in hand, not the whole watch. BASE did not choose BEAM by analogy. It chose it by physics, because it is the only substrate that gives forever-running, negotiating agents their guarantees.
• The primitive, done right (self-similar emergence). Get one agent correct, then let it interact with itself at higher orders: agent → cluster → fractal → holarchy. The four-domain table (BASE / Game of Life / biology / society) shows the same ladder — primitive, interaction, emergence, scale.
• What an agent is. A persistent entity with an identity and four sectors — Biography (what it did), Attributes (what it knows), Schedule (what it intends), Execution (what it is doing) — plus a cortex that keeps running even when no model is called. Checkpointed to disk; resumes from known state after any crash. The agent is not the conversation.
• How they will negotiate and settle — honest about what is built. Two things are already real and running: the annealing scheduler that settles each agent's own work toward a low-energy plan (Beat 3), and the peer-to-peer wiring that lets agents talk to one another directly. The contract-net negotiation between agents — the open market of Beat 2 — rides on that wiring and is the next thing I build. I say that plainly: the substrate is here; the negotiation layer is months away, not shipped.
• How it is allowed to improve. BASE does not impose a learning loop; it permits one, cleanly. Because every agent already keeps a Biography of what it did and an Attributes store of what it knows, you can write a behaviour that reads its own history, distils a lesson, and updates what the agent knows — a Plan-Do-Check-Act loop that is yours to author, not a policy I force on you. The architecture makes the loop natural. It does not make it mandatory. That is the difference between a platform and a straitjacket.
• Proven where software goes to die — and then I made a hard call. For three years we focused on building BASE systems for underground mining: no reliable network, no clean power, disconnected networks, and multiple different systems to integrate and organise. Technically it was a superb fit the architecture held in the worst conditions there are. But mining is a brutal industry for a startup, and in the end it was the politics, not the technology, that slowed us down. So I made a deliberate choice: rather than sink a decade into one hard contract and watch BASE drift into irrelevance while the world's tooling raced ahead, I pulled out to build the standard while the window is open. The mining years are not a customer logo. They are my proof that this architecture survives reality and the reason I know exactly how fast this field is now moving.