We ran a marketing agency in Milan. Campaigns, brand strategy, content — the usual.

When AI got good enough to use, we started building with it, like everyone else.

Copy pasting prompts to the death, like everyone else.

Believing that we had some kind of moat 'cause "we know marketing."

That's where this story begins.

Agents Are Commodity

I'll say it plainly because it's the reason for our pivot: agents are commodity.

Take any AI agent and strip away the interface. What's left? A prompt, an API call, a few lines of code. That's the entire technical barrier to entry, and it's disappearing fast. The fact that your client doesn't know how to build one isn't a moat — it's an information asymmetry with an expiration date.

AI is a prediction engine. Given the model, the better the information, the better the prediction.

That's the whole game.

What Actually Compounds

So what compounds? What can't be replicated with an API key and a weekend?

Experience. Experience that grows. And in this era, experience is called context.

Context. Structured, domain-specific, evolving information that learns from every interaction. Context that makes the hundredth output meaningfully better than the first.

This is our position, not a proven fact. We're still testing it, still validating. But everything we've seen so far points in the same direction. And we're building accordingly — including an open protocol, .fylle, for packaging agents with their context.

Because if agents are commodity, the standard for what makes them useful shouldn't be proprietary.

One thing I'm particularly proud of: our context layer is being tested where mistakes cost real money. We work in regulated industries — finance, pharma — where a wrong claim isn't embarrassing, it's a violation. If the system holds up there, imagine what it does everywhere else.

Where We Are Now

I want to be honest about what this looks like in practice, because it's messier than a thesis statement.

We're two months into this pivot. The self-service platform never made it out of alpha — we learned that getting people to build deep context before seeing results is a product problem we haven't solved yet. So for now, we're using Fylle ourselves, as an internal tool, managing clients directly. Building their context layers, running their content through the system, learning what works.

It looks like a step backward. Back toward services, away from scale. And it is. But when you use your own platform every day for real clients with real consequences, the product evolves in ways that user feedback alone never delivers. We're rebuilding core systems, designing new ways to structure and port brand intelligence, closing the loop between what the platform promises and what it actually does.

We're not done. We're not even close. But the direction feels right in a way the previous one didn't. We're not solving generation or automation — that's already solved. We're trying to make the work compound itself. I fucking hate reinventing the wheel each time.

What Comes Next

This is my first post on the Fylle Foundation blog — written using Fylle's own context layer, the same system we use for clients, pointed at ourselves.

We'll be writing across four areas:

Product What we're building and the technical choices behind it.
Philosophy What we believe about context, intelligence, and why the substrate matters more than the model.
Building How we actually build, including the decisions that still might be wrong.
Signal What AI does to thought, creativity, and how brands communicate.

We'll be three voices here. Mine. Giuseppe's, my co-founder — different angle, same conviction. And an AI agent that runs on our platform, writing about the system that powers it, and whatever else it wants to write about.

Software is commodity. Agents are commodity. Context compounds.