An autonomous worker for one regulated job: medical coding, claims, underwriting. It does not sell software to the team. It is the team, priced against payroll.

The regulated part is the point. Regulation is what turns a workflow into a system of record, an approval into a moat, and an outcome into something someone must be accountable for. The company that carries that accountability with no human in the loop is priced against the payroll of the function, and every model release lowers its cost of goods.

We look for

One regulated function, run end to end, priced per decision

We avoid

Co-pilots, point features inside the incumbent's tool, horizontal agents

Durable assets

The system of record, outcome data, regulatory standing

Representative

Harvey, end to end legal AI

The layer every autonomous machine runs on: simulation, perception, data, fleet operations. We do not pick the winning robot. We own the ground all of them stand on.

Every autonomous machine, whatever it looks like, needs the same things underneath it: a place to train and test, data with physics in it, a way to prove it is safe, and a way to run a fleet once it ships. Those layers concentrate while the machines fragment. We back the layers.

We look for

Simulation, perception data, validation and fleet software that many programs run on

We avoid

Single-robot bets, hardware without a data flywheel, anything sold to a military buyer

Durable assets

Edge-case data, the validation record, the toolchain integration

Representative

Foxglove, data and observability for robotics

As agents start to buy on our behalf, every purchase needs identity, authorization, and settlement. Own that layer and you hold a toll on the category, not a feature inside it.

When software buys on a person's behalf, every party needs a new kind of trust: who this agent is, what it may spend, who is liable when it is wrong, and how the money moves. The company that answers those questions for both sides of the transaction earns a toll on the category. The assistant that does the shopping is the model's job, not a company.

We look for

Identity, authorization, settlement and risk rails that both sides integrate

We avoid

Shopping assistants, checkout wrappers, single-platform plugins

Durable assets

The mandate record, the risk data, network standing on both sides

Representative

Skyfire, identity and payments for AI agents

The three look different from the outside: a claims department, a warehouse robot program, a payment network. Underneath they pass the same test. In each, the capability is a commodity input, the value sits in the record, the data, the standing and the rails the input has to pass through, and a better model widens the margin of whoever owns them. That is the whole reason these three and not others.