In the 1840s, Manchester, England did not suffer from a lack of invention. It suffered from too much of it.
Workshops across the city were turning out bolts, gears, shafts, couplings, each shop proud of its precision, each machinist confident in his own measurements. Steam engines multiplied. Railways stretched outward. Industry roared. And yet, beneath the smoke and progress, a quiet defect persisted.
Nothing quite fit.
A bolt cut in one factory might seize in another. A replacement part might rattle loose or refuse to turn. Thread angles varied. Depths varied. Turns per inch varied. Every shop spoke its own mechanical dialect. Manufacturing was advancing, but coordination was fragile. Industry remained local, even as it aspired to be global.
The Whitworth Moment
Into this mechanical Babel stepped Joseph Whitworth.
Whitworth did not build a more powerful engine. He did not invent a new machine. He proposed something almost embarrassingly simple: a standard thread, fifty five degrees, uniform depth, fixed turns per inch. A bolt forged in Manchester could now turn cleanly into a nut milled in Glasgow or repaired in London.
Whitworth did not improve the bolt. He standardized the interface between bolts.
And in doing so, he made fasteners fungible.
That move, modest as it seemed, altered the trajectory of industry. Factories could specialize because parts could travel. Supply chains could expand because compatibility no longer depended on personal familiarity. Innovation could compound because each new machine did not require reinventing the thread.
Whitworth did not improve the bolt. He standardized the interface.
Whitworth introduced a layer of compatibility. He did not change the machines. He changed the grammar by which they connected.
Fungibility in Computing
If you look closely at the history of computing, the same pattern repeats.
We tend to celebrate intelligence and novelty. We rarely celebrate interchangeability. And yet, progress in computing has advanced less by brilliance than by making one more thing reliably swappable.
Transistors made logic fungible. Operating systems made processes fungible. Internet protocols made hosts fungible. HTTP made messages fungible. The hyperlink made documents fungible. Each of these shifts introduced a common grammar that allowed independent systems to connect without renegotiating terms every time.
The web’s real breakthrough was not that it was clever. It was that a browser could approach an unfamiliar site, discover links and forms at runtime, and act without bespoke integration. HTML passed a simple test. Could we build a universal client for any system built on this grammar? We could. That was fungibility at work.
Over the last two decades, we have become exceptionally good at standardizing transport and syntax. We have refined protocols, tightened message formats, and generated increasingly precise contracts. We can describe the shape of a payload in exquisite detail. We can validate structure. We can enforce types.
What we have not standardized is semantic behavior.
Incompatible Semantics
Every API still defines its own meanings for account, order, member, subscription, approve, cancel, activate. These words are not decorative. They encode intention. They define what can be done and under what expectations. They are the threads by which systems coordinate.
In the age of AI agents, this gap has become easier to overlook.
Large language models sit on top of these bespoke systems and perform feats that would have seemed fanciful only a few years ago. They inspect schemas. They read tool descriptions written in prose. They choose actions. They compose sequences. In narrow domains, under careful supervision, they perform admirably.
It feels like coordination. Tt is closer to interpretation.
In previous eras, we reduced ambiguity before increasing abstraction. In the agent era, we increased abstraction without first reducing ambiguity. We introduced an engine of reasoning without first grounding it in standardized shared meaning.
An LLM does not encounter a shared semantic thread. It encounters text and patterns. It infers what a tool might do based on statistical regularities in language. It smooths over inconsistencies. It reconciles mismatched vocabularies through probability.
Guessing is not fungibility.
Whitworth did not solve Manchester’s problem by training machinists to become better at eyeballing mismatched threads. He did not invent a clever device that could adapt one arbitrary bolt to another. He fixed the angle.
The Missing Layer
We, by contrast, have introduced a powerful reasoning layer without first establishing a shared semantic layer beneath it. Our stack moves from transport to message to representation to contract and then leaps directly to reasoning. Between contract and reasoning lies an unformalized gap, the very place where shared meaning should be made durable.
This is not an argument against agents. It is an argument for completing the stack.
LLMs are extraordinary reasoning engines. They can plan, translate, summarize, and synthesize. They can operate as interpreters at the edge of heterogeneous systems. But reasoning is not grammar. Inference is not compatibility. A probabilistic bridge can span a gap for a while. It does not eliminate the gap.
Every durable computing revolution added a layer that made something interchangeable. None of those layers replaced the ones beneath. They constrained them, stabilized them, and allowed innovation to accelerate above them.
If we are serious about an agentic future, then we should ask a simple, unfashionable question. What, exactly, have we made fungible?
Messages are fungible. Documents are fungible. Hosts are fungible. Meaning is not.
The Next Whitworth Moment
And so we return to Manchester.
In 1841, it would have been possible to argue that each factory’s unique thread was a mark of craftsmanship. Standardization might have seemed like a threat to creativity. Instead, it became the condition for scale. Once bolts fit anywhere, industry ceased to be a collection of local dialects and became a coordinated system.
Today, machines can exchange messages across continents in milliseconds. They can draft plans and write code. They can converse fluently. Yet at the semantic level, they still speak in dialects. They exchange structure, not shared intention.
Messages are fungible. Meaning is not.
The next Whitworth moment in computing will not standardize transport. It will not standardize syntax. It will introduce a common thread for meaning in motion. Not by replacing agents, but by giving them something firmer than inference to stand on.
Whitworth did not make machines smarter. He made them compatible.
The agent age, like 19th-century Manchester, will not scale until the threads match.
