I need to tell you about something that happened to me last week.

I was at the Gartner Data and Analytics Summit. We spoke to over 200 companies in a few days. Fortune 500s, global logistics operators, financial institutions, major airlines. Serious people with serious budgets and real operations to run. And every single one of them had the same question. Not "how fast is it?" Not "what models do you use?" The question was: can I trust this with my business?

And then, at the end of one of these conversations, an executive at a major logistics company pulls me aside and goes, "So, sorry for asking, but... are you guys an agent?" And I said no. And he said, "Oh, thank God."

That's literally the first time that's happened to me since I started this company. And I think it means the market is catching up.

40% of agent projects won't make it. That's not a prediction. That's math.

Gartner predicted that over 40% of agentic AI projects will be canceled by 2027. Escalating costs, unclear business value, inadequate risk controls. They estimate that only about 130 of the thousands of vendors claiming to have agentic AI actually have anything real. The rest is what Gartner calls "agent washing": rebranding chatbots and RPA tools and slapping the word "agent" on them.

This isn't surprising to us. We've been watching this happen in real time.

We are not anti-agent. Agents are great for certain things. If you want to replace navigating UIs, if you want to make people a little smarter or a little faster at browsing the web or writing emails, go for it. Agents work when the stakes are low and the tolerance for error is high.

But here's where it breaks down. We work with an insurance company where a 0.5% error rate on claims processing would mean millions of dollars in liability. Millions. That's not a rounding error. That's a catastrophic business outcome. And agents today have massive error rates. They just do. If you're processing thousands of claims a day and your system is wrong even half a percent of the time, you're not saving money. You're creating a legal and financial disaster.

Access to inference should not require a conversation

We think the current way of interacting with AI, meaning chat-based APIs and turn-based conversation, is extremely inefficient and extremely unproductive. There are papers showing that agents lose accuracy the more turns you take on them. More turns, less accurate. That's the opposite of what you want from a system handling critical operations.

So we built something different. We call them functions. They're guardrailed, observable, promptless. Every function does a very specific thing. You can audit it. You can trace every decision back to the source. And you can chain these functions together to do increasingly complex things.

Instead of asking a chatbot whether you should approve a mortgage, you have a controlled, step-by-step pipeline where each step is so constrained that if the system makes a mistake, it would be immediately visible. An operator can flag it, and that mistake will never happen again. That's something you just cannot do with an agent today.

This is what we mean when we say we're the bricks. You bring the mortar. You bring the business logic, the policies, the workflows. We give you the primitives to build something that actually works in production, at scale, with the kind of accuracy that critical operations demand.

Programmatic SOPs, not vibes

Here's the concrete version of what this looks like. A customer comes in and says: these are my policies. This is how I triage invoices from vendors. If it's a large company, the invoice has to contain X, Y, and Z. It has to be under this amount, it has to match these criteria. We ingest all of that, in whatever format, because we already have the engine. We understand how your business operates. Your vendors start uploading invoices to get paid. We don't just extract the data. We make decisions on the data. We can help you underwrite loans, approve insurance claims, triage medical authorizations. And every single decision is completely auditable.

The system trains itself automatically based on your operator's feedback. Instead of your team reading every document and making every determination by hand, they supervise the work. They sit there and say, correct, correct, incorrect. And every time it gets something wrong, it feeds back into the system and that mistake is gone. Your team manages by exception rather than by the rule.

That is fundamentally different from an agent. An agent says "let me try this, and if it fails, it fails." Our system is controlled. Every step we make is constrained. We eliminate the variables so that it is very clear what happened and why, what the trace is, what the source is.

This paradigm didn't have a name. So we built one.

Under the hood, what we do looks mechanically similar to an agent. We're calling AI models multiple times. We're going back and forth. But every step is extremely controlled. Extremely guardrailed. And that distinction matters enormously when you're running a business where errors have real consequences.

We call them stochastic functions: a way to build products with statistical computing, with inference, without prompts. No one else in the market interacts with AI models this way. The rest of the industry is building on chat. We're building on controlled execution. That's not a subtle difference. It's a fundamentally different architecture for a fundamentally different set of problems.

The agent metaphor is pretty brilliant, I'll give it that. It anthropomorphizes AI. "A little person inside my computer that does things." People get it immediately. But getting it and trusting it with your business are two very different things. And the gap between understanding and trust is exactly where we operate.

The market is already telling us something

I spent a week talking to enterprise leaders and every conversation came back to the same thing. Is this accurate? Am I going to get fired for this? Is this something I can trust our business with? And the honest answer, for most agent platforms out there, is no. Not yet. Not for critical operations.

Our answer is yes. Because we are so conservative in the way we handle AI, in the way we call it, in the way we orchestrate it, that if there is a mistake, it would be immediately catchable. And every time there is a mistake, we take it out. Next time, that mistake won't happen.

So why are we not an agent?

Because our customers don't need us to be one. They need us to be accurate. They need us to be observable. They need every decision to be traceable. They need a system that gets better every time it makes a mistake, not one that might make the same mistake a thousand more times before anyone notices.

We're the bricks. We're the foundation. We're the thing you build on when the stakes are too high for vibes.

And honestly? I think that executive said it best when he looked me in the eye and said, "Thank God you're not an agent."

We couldn't agree more.

The painting above is The Red Horseman, by Roy Lichtenstein. The original lives at the MUMOK in Vienna. Lichtenstein explored futurism in the 1970s (what today we call "pop-futurism"), and he based this painting on the original by Carlo Carrà, a 1910s Italian futurist. With this rendition, he was a bit cheeky; while it looks like a homage, he detaches himself from the politics of futurism by "freezing" the horse instead of giving it movement.