The script for a technology wave never changes, and I have now watched five of them run it. Big data. Mobile-first. Microservices. Cloud. Now AI agents. Each arrived as a genuine shift wrapped in a hype cycle so familiar I could recite it before the conference keynote started.
The hype cycle runs in the same order every time. A real capability appears, then gets overclaimed by an order of magnitude. A wave of teams adopt it because their board read the same article, not because they have the problem it solves, while a smaller set of teams use it where it actually fits and gets a durable edge. Eighteen months later everyone agrees it was always obvious, and the same people who oversold it move on to oversell the next thing.
I have been on every side of that cycle, including the quiet winner. Early on I was also the engineer who rebuilt a perfectly good monolith into microservices because a slide deck told me to, and spent the next year discovering I had turned one debuggable system into forty distributed failure modes with a service mesh on top. That one left a scar, and it became the most useful thing the hype cycle ever gave me.
The same five mistakes, in a new costume
The waves are not that different from each other in how they go wrong. The technology changes completely, but the failure modes are reruns.
With big data it was “collect everything, the insight will emerge,” except it never did. We built petabyte lakes that became petabyte swamps, and the teams that won were the ones who knew which three questions they were actually trying to answer before they bought a single node. With microservices it was “decouple everything,” and the bill for that arrived as latency, eventual consistency bugs, and an on-call rotation that hated me. Cloud sold “lift and shift, it’ll be cheaper,” until the invoice taught a generation of us that elasticity you do not govern is just a faster way to set money on fire.
Every wave promised the boring middle would disappear, and it never did. The boring middle is where the value actually lives, and the hype is designed to make you forget that.
What I look at now instead of the demo
When AI agents showed up promising to do the work for me, I did not feel wonder. I felt the muscle memory of someone who has been lied to by a demo before, and I started running the same four checks I run on every wave.
The first check separates the durable capability from the narrative. Strip the story away and ask what the technology can genuinely do that nothing could do last year. For agents the honest answer is real and large. A model can now read a messy instruction, plan a few steps, call a tool, look at the result, and adjust, which is a genuinely new primitive. It is not what the narrative calls “an autonomous digital employee,” and the gap between those two is where most of the money is currently being lost.
The second check looks for where it removes real toil, not where it does something impressive. Impressive and useful are different axes, and the hype cycle works by conflating them. The agent that writes a sonnet in the demo is impressive; the agent that takes a reconciliation analyst’s four hours of cross-referencing ledgers and turns it into a reviewed twenty-minute pass is useful. I have built the second kind, and though nobody in the room claps for it, it is the only kind that pays for itself.
The third check is to distrust the demo structurally, as a category. Every wave has its hero demo, always the happy path with the edge cases swept off-camera. The mobile demo never showed you the offline state. The microservices demo hid the cascading timeout. The agent demo never shows you the run where the model confidently calls the wrong tool with plausible arguments, and you have no idea until the numbers are wrong downstream. A demo is a sales artifact, so treat it as one.
The fourth check is the one I trust most, because it has never once failed me. Find the boring integration work and assume that is where the value leaks out. The plumbing is always where it goes. With agents in 2026 that plumbing is what it has always been: permissions, data freshness, audit trails, the fallback when the tool call fails, the human checkpoint that has to add judgment instead of just adding latency. The model itself is the cheap part, the way it has been on every wave, and I have said so each time without once being wrong. What costs you is making that model trustworthy enough that someone will let it touch something that matters.
This is not cynicism, and the distinction matters
I want to be careful here, because pattern-matching the hype is one step away from becoming the bitter old engineer who thinks nothing is real and gets left behind. That person is also wrong, and more dangerously wrong than the hype-chaser, because the hype-chaser at least ships sometimes.
AI is real. The capability underneath the agent narrative is the most significant new primitive I have worked with in eighteen years, a bigger deal than any wave I have lived through, and I do not say that lightly after the cloud transition rewired an entire industry’s economics. I left a comfortable director seat to build on it, which is not what a cynic does.
The discipline is holding two thoughts at once: the capability is real, and the hype is still hype. The entire job of a technical leader during a wave is to keep those two facts from collapsing into each other. Collapse them one way and you ship a chatbot nobody asked for to please your board. Collapse them the other way and you are the person who said the internet was a fad.
What I am actually betting on is not the autonomous agent that runs the business, but the unglamorous layer underneath it: the integration fabric, the audit trail, the evals, the permission model, the checkpoints where a human still adds judgment. Call it the boring middle, the one every wave promised would vanish and never has. That layer is hard, it compounds, and the demo will never show it to you.
Eighteen years in, the technology keeps changing and my edge keeps coming from the same place: not from chasing the wave faster, but from being harder to fool about where the value actually went.