AI Agents: The Second Act

There is a lazy version of the comparison everyone is making right now, and a useful one. The lazy version says the agent boom looks like the dot-com bubble, so a crash is coming, so be careful. It is the kind of thing that sounds wise at a dinner party and tells you nothing you can act on. The useful version starts from a different place. It treats the dot-com era not as the lesson but as a single example of a much older pattern, and once you see the pattern, the future stops looking like a warning and starts looking like a map.

Here is the pattern, in one sentence. Every general-purpose technology gets adopted twice. The first adoption imitates the world that came before it, and that is the boom. The second adoption rebuilds that world around what the technology uniquely allows, and that is the payoff. The gap between the two is where the money is lost, then made, and it is almost always longer than the people living through it believe.

The web's first act was the brochure website. Its second act was Google, Amazon, and Salesforce. Agents are in their first act now. The single agent bolted onto a workflow is the brochure website. The decision intelligence workflow system is the Salesforce. Everything else in this essay is an attempt to earn that claim, using two hundred years of evidence and a 2026 report that is quietly confirming it in real time.

01 / The Law Beneath the Boom

This has happened before, on schedule

The economist Carlota Perez spent a career studying technological revolutions and found that they all move through the same shape. First comes installation: capital floods in, infrastructure gets built far faster than anyone can use it, and a financial bubble inflates on the promise of what the technology might one day do. Then comes a crash, which Perez treats not as the end of the story but as its turning point. Then comes deployment: the long, productive golden age when the technology actually reorganizes how people live and work. Installation is the frenzy. Deployment is the payoff. The crash is the hinge between them.

You can watch it run in the 1840s, long before software. Britain's railway mania saw hundreds of railway companies float on the promise of a connected island. A financier named George Hudson became known as "the Railway King," and then, predictably, the bubble burst and a great many investors were ruined. But the rails stayed in the ground. They became the circulatory system of the Victorian economy, and the people who built the durable businesses on top of them did so after the mania, not during it.

Two thinkers compressed this into aphorisms worth keeping. Roy Amara, of the Institute for the Future, gave us what is now called Amara's Law: we overestimate a technology's impact in the short run and underestimate it in the long run. Bill Gates said a similar thing in plainer words, warning that we overestimate the change coming in two years and underestimate the change coming in ten. Both are describing the same gap between installation and deployment. Both are describing exactly where agents sit today.

02 / The Web, Twice

From the sock puppet to Salesforce

Run the web through the pattern and every phase has a face you remember.

The land grab was "your business, but with a website." Between 1995 and 2000, the entire economy concluded that it needed a web page, and most of those pages were brochures: the old business, photographed and posted online, with nothing about the underlying work actually changed. It felt like transformation. It was mostly translation.

Then came the shakeout, and it was brutal and clarifying in equal measure. The Nasdaq lost roughly three quarters of its value. Pets.com, with its beloved sock puppet mascot, went from a Super Bowl ad to liquidation inside a single year. Webvan raised and incinerated a fortune trying to deliver groceries before the substrate could support it. Boo.com burned through more than a hundred million dollars in about eighteen months. The crash was not the failure of the web. It was the failure of businesses that had a web page but no durable job that the web made newly possible.

What the boom left behind was the substrate. Telecom companies had laid staggering amounts of fiber, far more than the traffic of 2001 could ever use, and only a small fraction of it was lit. For years that looked like the most wasteful overbuilding in corporate history. Then streaming, cloud, and the broadband internet arrived and lit it up. The bubble's most embarrassing excess turned out to be the foundation of the next decade.

And only then, on top of cheap overbuilt infrastructure, came the killer applications. Not brochures. Workflow-grade software that did things no prior process could. Amazon, which Barron's had mocked on a 1999 cover as "Amazon.bomb" and whose stock fell more than ninety percent in the crash, survived to rewire retail. Cisco, for one giddy moment the most valuable company on earth, fell hard, but the networking it sold became plumbing. Salesforce arrived preaching the end of installed software and proved that the enterprise would run on services delivered over the web. By 2011, Marc Andreessen could write that "software is eating the world," and he was simply narrating the deployment phase out loud.

The crash is not the failure of the technology. It is the failure of the businesses that imitated the old world instead of rebuilding it.

03 / The Thirty-Year Lesson

Why electricity took forty years to pay off

If the web shows the shape of the cycle, the electric dynamo explains why the gap between installation and deployment is so long, and what finally closes it. It is the single most useful precedent in all of this, and almost nobody invokes it.

Factories had access to electric power from the 1880s. The productivity gains did not arrive until the 1920s. For roughly forty years the technology was present and the payoff was absent, which is the kind of thing that makes economists nervous. In 1987, watching computers spread without moving the productivity needle, Robert Solow delivered the era's most famous lament: "You can see the computer age everywhere but in the productivity statistics." Nearly forty years later, in early 2026, an economist named Torsten Slok updated it almost word for word, observing that "AI is everywhere except in the incoming macroeconomic data." The paradox is a recurring character.

The economic historian Paul David explained why, using the factory floor. The first factories to electrify used the new power exactly as they had used steam: one enormous central motor driving a maze of shafts and belts that ran the whole building. They swapped the engine and kept the architecture, and so they got almost nothing. The gains came only when a later generation threw out the central driveshaft entirely and gave every machine its own small motor. That one change freed the floor plan, let machines be arranged around the logic of the work instead of the location of the power, and culminated in Henry Ford's moving assembly line in 1913. The payoff did not come from electrifying the old factory. It came from designing a new one.

The payoff never came from electrifying the old factory. It came from designing a new one around what electricity made possible.

This is the lesson the agent boom is fumbling in real time. Bolting an agent onto today's process is the central driveshaft with an electric motor strapped to it. It runs, a little better than before, and it convinces no one. The revolution is the workflow redesigned around what agents uniquely allow, built by the people who stop trying to automate the old decision and start rebuilding the decision itself.

04 / The Agent Boom on the Map

We are standing in the first box

Lay the agent timeline directly beneath the web timeline and the phases line up beat for beat.

The agent land grab is "your business, but with an agent." Every company now feels it must ship one, and most of what ships is the old process with a chatbot stapled to the front. There is no sharper proof that the cycle is repeating than watching the very companies that won the last one line up to imitate the old world again, this time with agents.

Consider Salesforce. It was the web's deployment-phase trophy, the company that proved enterprise software could live in the cloud. Twenty-five years later its founder, Marc Benioff, has turned the entire company toward agents, declaring that "software is about to become digital labor" and that he expects to spend the rest of his tenure managing human and digital workers side by side. Whatever you make of the marketing, the structural signal is unmistakable. The killer app of the previous cycle has become a land-grab participant in the new one. The wheel is turning, and the people who rode it last time can feel it.

The shakeout has not arrived yet, but the cracks are already visible to anyone reading carefully. Gartner's 2026 Magic Quadrant for AI Platforms for Data Science and Machine Learning is, when you read its cautions rather than its rankings, a catalog of fragile scaffolding. Almost none of the warnings are about the intelligence of the agents. They are about the structure around the agents: weak context layers, fragmented orchestration, immature governance, and unpredictable cost. That is precisely the profile of an installation phase nearing its reckoning. The agents demo beautifully. The things that would let them survive in production are the things the report keeps flagging as missing.

05 / What Survives the Crash

The model is the electricity, not the factory

Here the analogy strains, and the strain is the most valuable part of the argument, so it is worth meeting head on. The web's substrate was durable in the most literal sense. Fiber laid in 1999 still carries traffic today. The frontier AI model has no such permanence. It is overtaken roughly every eighteen months by something cheaper and stronger. A skeptic stops here and says the comparison collapses, because the core technology of this revolution is perishable in a way fiber never was.

They have it backward. The model's short shelf life is not a flaw 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 electricity: essential, ambient, and substitutable. You would no more build your durable advantage on a specific model than a 1920s manufacturer would have built his on a particular brand of motor.

What survives the crash is everything the model plugs into. The context and data layers that encode what a business actually means by its own terms. The protocols that let agents talk to each other and to the systems that run a company. The rails that move work safely between those systems. And the hard, unglamorous organizational fluency of running autonomy without losing control. These compound over time and do not reset when the next model ships. You can swap the engine in an afternoon. You cannot rebuild years of accumulated context, or an organization's muscle memory, with an API call.

This is exactly what the Gartner cautions are pointing at, even though the report never frames it this way. When it warns that a platform lacks a context layer, or struggles with orchestration, or has thin governance, it is naming the durable substrate by listing its absence. The market is, in its own bureaucratic language, already telling you where the moat will be. It will not be the model. It will be the wiring around it.

06 / The Second Act

Decision intelligence is the deployment-phase product

Every adoption's second act produces a signature artifact, the thing the deployment phase was actually for. The railways gave us the national market. Electricity gave us mass production. The web gave us the workflow application that runs the modern enterprise. In each case the signature product was not a faster version of the old thing. It was a new thing that the old world could not have held.

For agents, the shape of that signature product is already legible in the language the field is reaching for. The most telling phrase in the entire 2026 Gartner report describes the frontier as a shift "from isolated agents to interconnected decision workflows." Read that again, because it is the whole future in six words. Not a smarter assistant doing a task. Many agents coordinated into a decision that crosses systems, carries context, and stays governed from beginning to end. Gartner even attaches a number to the trajectory: it expects at least 15 percent of day-to-day work decisions to be made autonomously through agentic AI by 2028, up from essentially none in 2024.

Call it decision intelligence. The exact name will settle later, the way "e-commerce" and "SaaS" only settled once the things they named already existed and needed a label. What matters now is the architecture, and the architecture is the dynamo's lesson applied to autonomy. Stop strapping agents to the old decision. Start designing the decision around what coordinated agents make newly possible. The winners of this cycle will not be the people with the best agent. They will be the people who rebuilt a decision that used to require a room full of humans, many systems, and a week of back and forth, into something a coordinated workflow can carry end to end.

The single agent is the brochure website. The decision intelligence workflow system is the Salesforce. We just have not met it yet.

The boom is loud right now, the way every installation phase is loud, because installation is the phase that makes the most noise and the least lasting product. The quieter and more valuable question is the one this pattern has been posing for two hundred years. Not what can the new tool imitate. What can finally be rebuilt.