HAT over MASS

Like any company — and startups perhaps most of all — we live and breathe AI every day, riding the wave of innovation this new age has brought.

Our own journey in AI has been both exciting and difficult: exciting, because the potential is plainly enormous; difficult, because for a long time we could not find a way to enjoy that upside without also accepting the downside.

For a financial firm, the dilemma bites hard. Unleash AI's full potential and you inherit its risks; insist on full control and you cap the very potential you came for. We think we have found something that works — at least for us, at least in our industry.

We have a name for it: HAT over MASS — two concepts, each with two indispensable sides.

The first concept is HAT — shorthand for Human-Commanded, AI-Operated Trust. In finance, trust is essential for us to operate at all; and in the AI age, we have learned that trust has two elements, neither of which can substitute for the other. We must be able to trust that humans are fully in control of the system, and that AI is running at full potential, not throttled into harmlessness. A system with only the first is a straitjacket; only the second, a gamble.

The second concept is MASS — shorthand for Micro Agents, Short & Simple tasks. It matters because it is the mechanism that delivers both kinds of trust at once. The word carries two meanings, and we mean both. The first MASS is the acronym: each agent kept small, its job brief and narrow enough to be verified completely, every time. The second MASS is the word itself: sheer numbers of such agents, working side by side.

The two halves lock together: smallness is what lets humans control the mass; scale is what lets AI reach its potential. That is the essence of HAT over MASS.

This is not just a theory. We are already running on it. Today we have created more than 5,000 AI employees in our system — each with an employee ID, each assigned a functional role, and all of them already on the job. And we will publish their first results on October 3, when we release our third white paper in as many months. We have learned much from others who shared their AI journeys publicly; in the same spirit, we would like to share ours. Here is the arc: what we are building and the market we are after; why AI is existential to it; the attempts we made and failed at; the lessons we drew from them; and how we finally came across — or invented — the structure just described.

Here is how it all began.

Securitizing the roots of the economy

Our work starts from a conviction about where capital markets must go next. Wall Street's great invention was securitization — turning the cash flows of companies, mortgages, and loan portfolios into instruments the market can price and trade. The next chapter, we believe, takes that logic one level deeper: not securitizing at the company level, but at the contract level.

By a contract, we mean any atomic unit of economic cash flow. It could be the daily revenue of a brick-and-mortar store. It could equally be the cash flow of an algorithmic trading strategy, a stream of royalties, or the offtake agreement on a cargo of oil. What matters is not the size or the glamour of the underlying, but that it is a real, verifiable contract generating real cash. Distributed, miniaturized, atomized — financing the economy at its roots, through contracts rather than balance sheets.

Why the old model cannot follow us down there

Such a market makes brutal demands: origination and due diligence must be massively distributed, granular, real-time, high-frequency — disclosure not quarterly but daily; underwriting not of thousands of companies but of millions of contracts. On that terrain, Wall Street's high-cost model does not work. The economics of a banker do not scale down to the economics of a single contract.

Two conditions had to be met, and it is worth being precise about them.

A necessary condition, quietly fulfilled

The first has already been satisfied: the digital and payments revolutions have made the real economy transparent at the ground floor. Every sale leaves a data trail; every till is a sensor.

The sufficient condition we spent years looking for

The second, for years, we could not find. Transparency alone does not create a market. Someone — or something — must still originate, diligence, price, and monitor each contract, across millions of underlyings. No army of human analysts can do that; the cost structure defeats it before it begins. We became convinced, some years ago, that the missing piece would be artificial intelligence — and we were grateful when it arrived. With agentic AI, the sufficient condition finally seemed within reach.

Then we actually built with it

That is when we ran headlong into the dilemma this essay is really about: we wanted both kinds of trust at once, and nothing on the market delivered them.

The field, as we watched it, was splitting into two camps. At one end, some simply let AI run — potential first, control later. At the other end, where most of our industry sits, AI is admitted only as a tool, with a human approving everything it touches. Safe, but capped: if a person must check every move, you have not gained a worker; you have gained an expensive photocopier.

Our failed sandwich

We tried the middle road ourselves — what we privately called the sandwich: layers of AI and layers of people, the AI proposing, the people reviewing, round and round. It failed. It was slow, it was costly, and worst of all, accountability stayed blurry: when something went wrong, who owned it — the layer that proposed, or the layer meant to catch it? The sandwich did not answer the question; it merely made it more expensive.

For a while we sat with an uncomfortable thought: were we condemned to a false choice — potential without control, or control without potential?

The answer was already on the highway

The idea that freed us came from another field entirely — an inspiration, not a ready-made answer. Consider Tesla's Full Self-Driving: nobody approves each turn of the wheel; the car drives itself, at full speed, in real traffic — inside rules it did not write and cannot cross.

The humans have not disappeared; they have moved up — from steering to owning the rules. Every rule change is recorded; everything the car does can be replayed. The road gave us the inspiration; our own practice would have to provide the evidence.

Could we build that for knowledge work — trust that is human-commanded but AI-operated? Not human-checked (that was the sandwich) — human-ruled: the human writes the boundary; the AI runs free inside it.

Shrink the individual, grow the crowd

That was the first half of the answer —the hat. The second half — the mass —came from an observation about cost. One more AI agent costs almost nothing — roughly the price of one more software seat. So instead of asking one brilliant AI to carry out a long, complex task — which is precisely where today's models are weakest — we inverted the problem.

If the unit of finance shrinks to a single contract, the unit of work has to shrink with it. Keep every agent small: one short, simple, verifiable job — then deploy them by the thousand. Narrow the task, grow the crowd — the crowd can be enormous, because each addition costs next to nothing. What shrinks is the assignment, never the capability of the model behind it. Many small things, not one big thing.

The industry is shipping our building block

We recently felt a good deal less alone in this. A new kind of model has appeared on the market — Jev, from TypeSafe AI — built to do exactly one thing: render a small, well-defined judgment. You give it the material, a narrow question, and a fixed answer set — choose a category, yes or no, which grade — and it returns a structured, calibrated judgment, quickly and cheaply. No essay, no long chain of reasoning; a "System 1" for software, in their framing. We read that with some pleasure — it is exactly the job description of one of our agents. The industry, it seems, has independently arrived at the very building block we are made of: small, bounded, verifiable judgments, repeated at scale. The wind has started blowing the way we were already sailing.

Putting the hat on the mass

The rule each agent works under, we call a hat. It sits on the agent's head; a person put it there; a person can change it. Of any agent, at any moment, you may ask: what hat are you wearing, who gave it to you, when was it last changed?

That is the essence of the structure: the human manages the hat; the agent does the work. Accountability does not dissolve into the crowd — it keeps a name and an address. A hat for every agent; a mass of agents for the firm; the hat over the mass.

A structure that shows up for work every morning

And the part we are most glad to report: the structure is already running inside our firm, around the clock. We spare you the details — they deserve their own telling — but these reflections come from practice, not from a whiteboard.

What has it given us? The two kinds of trust we thought we had to choose between. The crowd moves at full speed; every agent operates inside its hat, so the control is real — written, recorded, replayable. And trust, we have learned, is not earned once but re-earned daily. There is no savings account for trust; we re-prove it every day, and our ledger is open.

Every generation of tools bequeaths an artifact. The abacus gave us the ledger book. The spreadsheet gave us the cell. SaaS gave us the database row. What AI leaves behind — what the daily cash flow of every contract becomes inside such a system — is something we will show, not tell.

Why we are telling you this

We share this because many of you may be standing at the same fork where we stood. We do not claim to have figured everything out — only that after many failed attempts, we found one structure that works, and more than 5,000 AI employees who are now working around the clock every day to prove it.

Five thousand is not a round number we reached for. Every agent sits in one of four functions —Administration, Business, Control, and Deployment. Consider a single task: generating the industry compass for each of 500 verticals. In every vertical we run nine agents — three viewpoints (the good-faith operator, the seasoned insider, the adversary), each on three different foundation models — three by three, nine per vertical. Add the shared Administration, Control, and Deployment seats across all verticals, and the total lands precisely at 5,000. Each ID stays with its holder across tasks; each hat is owned by a named person whose accountability travels with them through reassignments; every call leaves a ledger entry. More will come, but no agent joins the ranks until a named human is ready to answer for it.

Each month, on the 3rd, we publish one; October 3 brings the third — HAT over MASS itself: how the structure performs in real practice, with the 5,000 on the job. In it, we lay out our methodology for building Industry compasses out of the cash flows of individual contracts, and put all 500 vertical compasses on the table for market scrutiny.

Today, we have shared only the theory. On October 3, we will show you the machinery — how 5,000 AI employees actually clock in, do the work, and stand up to scrutiny, in the real-world arenas of small and micro finance. Not a plan, but a system already at work.