Everyone reaches for the dot-com comparison when they talk about the agent boom. The valuations, the hype, the coming crash. It is an easy comparison and a shallow one, because it treats the web as the lesson. The web was not the lesson. The web was one data point in a pattern that has repeated through every technological revolution for two centuries. Railways did it. Electricity did it. The automobile did it. The web did it. Agents are doing it now.
The pattern is simple enough to state in a sentence: every general-purpose technology is adopted twice. The first adoption imitates the world that came before it. The second adoption rebuilds that world from scratch. The gap between the two is where fortunes are made and lost, and it is almost always longer than the people living through it expect.
This is not an argument about whether agents are overhyped. They are, in the precise and temporary way every important technology is overhyped at this stage. This is an argument about what comes after the hype burns off, because the pattern tells us, with unusual clarity, exactly what that is.
Installation, then deployment
The economist Carlota Perez spent a career studying technological revolutions and found they all move through the same two phases, separated by a crash.
The first phase is installation. New technology arrives, capital floods in, infrastructure gets built far faster than anyone can use it productively, and a financial bubble inflates on the promise of what the technology might one day do. This phase ends in a crash. The crash is not the failure of the technology. It is the failure of the speculation wrapped around it. The railways still ran after the railway mania collapsed. The fiber stayed in the ground after the dot-com crash.
The second phase is deployment. On top of the cheap, overbuilt infrastructure the bubble left behind, the real applications get built. This is the productive golden age, the period when the technology actually changes how people live and work. The web's deployment phase produced the company that organized the world's information, the company that rewired retail, and the entire category of software that now runs the enterprise. None of those were brochure websites. They were built after the brochure websites died.
Read against this map, the agent boom is unmistakably in installation. Capital is flooding in. Inference capacity is being built far ahead of productive demand. And the dominant product is the agent bolted onto an existing workflow, which is this era's brochure website: a thin layer of new technology laid over a process that has not changed.
The thirty-year lesson of the dynamo
If installation and deployment describe the shape of the cycle, the electric dynamo explains why the gap between them is so long, and what finally closes it.
Factories electrified slowly. Electric motors were available for decades before they delivered the productivity gains economists expected, and for a long stretch the gains simply did not appear. The reason was architectural. The first factories to adopt electric power used it the way they had used steam: one giant central motor driving a system of shafts and belts that ran the entire building. They swapped the engine and kept the factory.
The gains arrived only when a later generation rebuilt the factory around what electricity uniquely allowed. Distributed power meant every machine could have its own motor. That freed the floor plan from the tyranny of the central driveshaft, let machines be arranged around the logic of the work rather than the logic of the power source, and eventually produced the assembly line. The payoff did not come from electrifying the old factory. It came from designing a new one.
This is the most important thing the pattern teaches, and the thing the agent boom is getting wrong in real time. Bolting an agent onto today's process is the central driveshaft with an electric motor strapped to it. It works, marginally. It is not the revolution. The revolution is the workflow designed around what agents uniquely allow, built by people who stopped trying to automate the old decision and started rebuilding the decision itself.
The agent boom, phase by phase
Place the agent timeline directly beneath the web timeline and the phases line up beat for beat.
Land grab. The web's land grab was "your business, but with a website." The agent land grab is "your business, but with an agent." Both are the new tool imitating the old world. Both feel like transformation and are mostly translation.
Shakeout. The dot-com crash cleared out the companies that had a website but no durable job to do. The agent shakeout is coming for the same reason: the wrapper that demos beautifully and collapses in production. The early signs are already visible in how the market evaluates these platforms. The hard questions have stopped being about the agent's intelligence. They are about everything around it.
Substrate. The web boom's lasting gift was the infrastructure it overbuilt: fiber, broadband, payment rails, and a generation that learned to live online. The agent boom is laying down its own substrate right now: inference capacity, the data and context layers that make autonomy useful, interoperability protocols between agents, and an organizational muscle for supervising things that act on their own.
Killer application. Built last, on top of the cheap substrate, by the survivors who chased a durable decision instead of a demo.
What survives the crash
Here the web analogy strains, and the strain turns out to be the most useful part of the whole exercise.
The web's substrate was durable in the most literal sense. Fiber laid in 1999 still carries traffic today. The frontier AI model does not have that property. It depreciates in roughly eighteen months, overtaken by the next one. A skeptic will stop here and say the analogy breaks, because the core technology of this revolution is perishable in a way fiber never was.
They have it exactly backward. The model's short shelf life is not a hole in the argument. It is the argument. If the most visible and most expensive component of the system is also the one that goes obsolete fastest, then the model cannot be the thing that lasts. It cannot be the moat. It is the perishable input, the electricity, not the factory.
What survives the crash is everything the model plugs into: the rails that move work between systems, the protocols that let agents coordinate, the context and data layers that encode what a business actually means, and the institutional knowledge of how to run autonomy safely. That is the durable substrate of this revolution, and it is the same set of capabilities the market is already learning to value over raw model performance.
Decision intelligence is the deployment-phase product
Every adoption's second act has a signature product, the thing the deployment phase is actually for. For the railways it was the national market. For electricity it was mass production. For the web it was the workflow application that runs the enterprise.
For agents, the shape of the second-act product is already visible in the language the field is starting to use: not agents, but decision systems. Not a clever assistant doing a task, but many agents coordinated into a decision that crosses systems, carries context, and stays governed end to end. The unit of value moves up, from the single autonomous actor to the orchestrated workflow that produces a trustworthy outcome.
Call it decision intelligence. The naming will settle later, the way "e-commerce" and "SaaS" settled only once the things they named already existed. What matters now is the architecture, and the architecture is the lesson of the dynamo applied to autonomy: stop attaching agents to the old decision, and start designing the decision around what coordinated agents make possible.
The boom is loud right now, the way every installation phase is loud. The quieter and more valuable question is the one the pattern has been asking for two hundred years. Not what can the new tool imitate, but what can finally be rebuilt.