Agentic-as-a-Service and the Return of the Engineer

I've been an engineer since the late nineties.

I started in electronics, because that's what you did back then if you wanted to work in media. To be a cameraman or a sound engineer you had to study something technical first — this was the era when a digital camera weighed thirty kilos and you carried it on your shoulder like a sack of cement. From there I drifted into computers: first fixing them, then, by 2005, building e-commerce when "having a shop online" was still a slightly exotic idea you had to sell people on.

I got in early. I still remember the first time my father set up a dial-up modem and a laptop at home in 1996 and let me play with it. It screeched, it dropped the connection if anyone picked up the phone, and it was, by any reasonable measure, an extremely expensive toy. But I was hooked. I had a front-row seat to the arrival of the internet, and I never gave the seat back.

Somewhere in my thirties, all of that turned into a strange and very useful skill: I could see where technology was going. Not vaguely — concretely, and with enough lead time to act on it. My edge, the thing I'd tell people over a beer, was that I could see roughly two years ahead. Two years was enough. Two years is the difference between building the thing everyone will need and building the thing everyone already has. I rode that instinct through startups and a fair share of entrepreneurial adventures, and it rarely let me down.

Then LLMs went mainstream, and my two-year horizon collapsed to two or three months.

I want to be honest about how that felt: it worried me. The skill I'd quietly relied on for a decade and a half stopped working overnight. Not because I'd lost it, but because the ground itself had started moving faster than anyone's intuition could track. When the state of the art resets every quarter, "where will this be in two years" stops being a forecast and starts being a coin toss.

The hype is finally burning off

Here's the thing about a craze: it can't last forever. And we're now at the point where it's starting to cool — not the technology, the noise around it. We're finally able to separate the media's hyper-inflated hype from the actual developments underneath, and the gap between the two turns out to be enormous.

The clearest signal is in the layoffs that were supposed to be the future. The companies that didn't fire 40% of their workforce in a fit of AI optimism can now quietly count themselves among the lucky ones. Klarna is the cautionary tale everyone points to: it shrank its headcount by around 40% and bragged that its OpenAI-built assistant was doing "the work of 700 agents" — and then, by 2025, reversed course and started rehiring humans after service quality fell off a cliff. Its own CEO admitted, in plain words, "we went too far." (https://www.reworked.co/employee-experience/klarna-claimed-ai-was-doing-the-work-of-700-people-now-its-rehiring/)

That's not an anti-AI story. It's an anti-hype story. The companies that treated the model as a magic wand got burned. The interesting question is what the companies who treat it as an engineering problem are about to build.

Agentic-as-a-Service (I am not writting the acronym to this...)

With the rise of what people are calling Agentic-as-a-Service — and I know, the name is not fortunate, please someone come up with a better name — we're finally getting our first real look at the next one-to-two years. And for once I feel like my old horizon is coming back into focus.

So let me make the bet plainly: as agentic systems and the harnesses around LLMs grow more complex, we will see an explosion of software services — the same explosion we saw with SaaS — just built on a different underlying technology. Not mass layoffs. The opposite: a ton of work for the engineers who master the new medium.

To see why, you have to look at what an "agent" actually is, because the marketing makes it sound like a personality and it's really an architecture.

A SaaS product is deterministic software you rent. You click a button, the same thing happens every time, and the vendor's job is to keep the lights on and ship features. An agentic service is something else: it's a system that performs work and delivers an outcome, and the model is only the smallest, cheapest part of it. The model is the engine. The car is everything else.

That "everything else" — the harness — is where the engineering lives:

• Orchestration and planning. Real tasks aren't a single prompt; they're loops. Decompose the goal, take a step, observe the result, decide the next step, recover when it goes sideways. That control flow is software, and it's hard software.
• Tools and integration. An agent that can't act is a chatbot. Giving it the ability to query a database, call an API, file a ticket, or move money means building, securing, and rate-limiting every one of those tool surfaces — and deciding what it's allowed to touch.
• Context and memory. Models have no memory between calls. Retrieval, state management, and long-term memory are entire subsystems that someone has to design so the agent knows what happened five minutes — or five sessions — ago.
• Verification and guardrails. This is the part the Klarna-style failures skipped. A stochastic system that's right 95% of the time is, in production, wrong one time in twenty — and that twentieth time is talking to your customer or touching your ledger. Checking the agent's work, constraining it, and knowing when to escalate to a human is non-negotiable engineering.
• Evals and observability. You cannot improve what you can't measure, and you can't measure a non-deterministic system with traditional tests. A whole new discipline — evaluation harnesses, regression suites for behavior, tracing every decision — has to be built around it.

Notice what just happened. Every bullet on that list is more work for engineers, not less. The model commoditizes the easy 80%; the differentiation — the moat — moves entirely into the system wrapped around it. And systems are built, debugged, monitored, and operated by people who know how to build systems. The non-determinism doesn't remove the engineer from the loop. It demands a better one, because we're now shipping software that's probabilistic, and reasoning about probabilistic systems under load is one of the harder things our profession does.

That's the shape of the next two years, as far as I can see it again: not fewer software companies, but a new generation of them — outcome-priced instead of seat-priced, agentic instead of static — and a deep, sustained demand for engineers who actually understand the medium.

What this doesn't mean

I'm not going to pretend the disruption is free. It isn't. Entire industries will transform, and some roles will have to admit, honestly, that they've become obsolete and will be swept out of the market. The traditional call center is the obvious one. Pretending otherwise helps no one, least of all the people in those roles who deserve a clear-eyed picture and a path forward.

But I have no interest in writing either an apology for the new or an elegy for the old. Both are lazy. The new isn't automatically good and the old wasn't automatically wise. My job — the actual work — is to find the paths forward: paths that run through better technology, and through the betterment of society as a whole, within my humble scope.

My two-year horizon is coming back. The view from here is busier, weirder, and more demanding than the hype merchants promised — and far more hopeful than the doom-mongers want you to believe. There's a tremendous amount to build.

I, for one, can't wait to get to work.