Agent labor is the new subscription

In December 2024, I wrote that SaaS applications would collapse into data stores as AI agents absorbed the business logic layer. That the intelligence tier was moving. That the seat was becoming irrelevant.

Fourteen months later, $800 billion disappeared from software stocks in a single week. The S&P 500 Software Index dropped 13% in five trading sessions. Salesforce is down nearly 30% year-to-date. Bloomberg coined a name for it: the SaaSpocalypse.

Everyone is writing the obituary. I'd rather write the birth announcement.

SaaS isn't dying because something went wrong. It's dying because something better arrived. License became subscription. Subscription is becoming agent labor. The pattern is the same. The panic is optional.

It has never, in the history of humanity, been cheaper to be a risk-taker in this game.

The market caught up

Public SaaS growth rates have declined every single quarter since the 2021 peak. Every quarter. For three years, the fundamentals were signaling what the stock prices refused to admit. The agentic AI narrative gave the market permission to finally reprice.

The iShares Expanded Tech-Software Sector ETF entered a technical bear market, down over 23% year-to-date. Piper Sandler downgraded Adobe, Freshworks, and Vertex on fears of seat compression. The pattern is consistent: if your revenue depends on the number of humans clicking buttons, your business model has a ceiling that gets lower every quarter.

When Satya Nadella said SaaS applications would collapse in the agent era, he was describing what builders had already been working against. The thesis from our December 2024 piece hasn't changed. AI agents absorb the business logic. The applications reduce to their barebones function: databases. The intelligence tier moves to the agent layer.

What changed is that $800 billion in market cap now agrees.

The business logic is all going to these AI agents. They're not going to discriminate between what the backend is.

Satya Nadella

The pattern is older than software

Nobody panics when you tell them the telegraph replaced the Pony Express. Nobody mourns hand-typeset printing. On a civilizational timescale, these transitions are unremarkable. Expected, even. Tools get better. Knowledge compounds. The delivery mechanism changes.

Perpetual licenses gave way to subscriptions. Everyone called it the death of software. It was an upgrade. Predictable revenue, continuous delivery, lower upfront cost. Better economics for both sides.

On-prem gave way to cloud. Same panic. Same result. The better infrastructure model won because it reduced operational overhead and scaled with demand rather than procurement cycles.

Now subscriptions are giving way to agent labor. The pattern is identical. The fear is identical. And the outcome will be identical: the model that delivers more value per unit of cost will win.

The SaaS model charged per human sitting in front of a screen. When an agent can screen a counterparty in minutes instead of four hours, the seat becomes a billing artifact from a previous era. You don't need the seat. You need the work done.

You should hope this is happening. If tools and knowledge stop evolving, we as a civilization are not moving.

What agent labor actually costs

Agent labor is a specific thing, not a buzzword. It means: software agents that autonomously perform defined tasks, produce auditable outputs, and escalate to humans only at decision points.

The SaaS model sold access. A dashboard. A workflow builder. A place for a human to do the work, slightly faster than with spreadsheets. Agent labor sells the work itself. The agent screens, classifies, gathers evidence, and delivers a structured decision. The human reviews and approves. That is the entire interaction.

This changes the billing model at a structural level. SaaS revenue tracked headcount. More employees, more seats, more revenue. When companies cut teams, the vendor lost revenue. The incentives were misaligned: the vendor needed the customer to keep hiring.

Agent labor revenue tracks work volume. More compliance reviews needed? More agent capacity consumed. When a company cuts its compliance team from ten to three, it needs more agent labor, not less. Revenue is tied to the work the business actually needs done, not to the number of humans doing it.

And agent labor is not free. We run production agents daily. A working agent that produces code, screens data, and executes real workflows costs $200 to $300 per day in API and compute. That is $75,000 to $110,000 per year. Real money.

But cost is not the real advantage. Velocity is.

A senior compliance officer earning roughly 120,000 euros per year doesn't lose on price. They lose on speed. A screening that takes four hours takes an agent minutes. A regulatory evidence pack that takes a week takes an agent a day. The gap isn't cost. The gap is that the agent-powered company makes decisions faster, responds to regulators faster, onboards clients faster. Compound that across a quarter and the competitive distance becomes unrecoverable.

This is not theory. We run fully autonomous agent workflows in production at customer sites today. Not pilots. Not demos. Shipped systems that screen, classify, gather evidence, and deliver structured decisions. The human reviews and approves. That's it.

None of my engineers have written code manually since December 2025. We're an AI development studio. Code was our main ingredient. Now our engineers do what they were always meant to do: architect systems, make judgment calls, solve problems that require understanding the domain. The agents write the code.

Code was our main ingredient for decades. Now it's a byproduct. The ingredient is knowing what to build and why.

Frederik Bonde, founder paterhn

We build faster. We build more consistently. And the customer gets a better system because the human attention goes to design and decisions, not syntax.

I expect code itself to be fully solved before this year is out. That's not a prediction I'm hedging. If I'm wrong, I'll say so. But every week the evidence gets harder to argue with.

The economics of work are being repriced. But the winners won't be the cheapest. They'll be the fastest.

The infrastructure layer is consolidating. Build above it.

This month, the creator of OpenClaw, the most popular open-source AI agent in the world, gave it away and joined OpenAI. His reasoning: building a company around it "isn't really exciting." The product wasn't a business. It was plumbing.

OpenAI didn't acquire a SaaS product. They hired the person building the agent infrastructure layer. They want to own the runtime where agents live. Not the layer where humans click buttons.

Within hours, the competitive retaliation started between the largest AI labs. Access revoked. Terms of service rewritten overnight. The infrastructure layer is now an active warzone, and if you're building at the plumbing level, you're building on someone else's floor. The landlord can change the locks.

This is why the value that compounds lives above the infrastructure. Domain expertise. Regulatory intelligence. Operational patterns specific to an industry, a jurisdiction, a workflow. The infrastructure will commoditize. It always does. TCP/IP commoditized. Cloud compute commoditized. Foundation models are commoditizing right now.

What survives every platform transition is the knowledge encoded in systems that understand the actual work. The hidden cost of building on someone else's model isn't the API bill. It's the dependency. It's the fact that your competitive advantage lives in a system you don't control.

The builders who survived the license-to-subscription transition weren't the ones clinging to the old model. They were the ones who understood the underlying work well enough to deliver it in the new one.

Same pattern. Same opportunity. And right now, the window is wide open.

The tools are available. The models are capable. The cost is known. The only question left is how fast you move. Because every week you're still paying per seat for work an agent could be doing, a competitor somewhere is compounding velocity you can't buy back later.

The panic is optional. The building is not.

The SaaS era rewarded companies that bought the right tools. The agent era rewards companies that own the right intelligence. The difference is whether the knowledge your business runs on lives in a vendor's model or in a system you control.

When you're ready to scope what agent labor looks like for your specific workflows, talk to an engineer. We've been building this since before the market noticed.