How small can a real company be now?

Starting a company used to mean assembling people before you could assemble anything else. Someone to chase the pipeline, someone to keep the books, someone to watch the market, someone to keep the project plan honest. None of that was the idea you started with; all of it was the price of operating. That price just collapsed. A single founder, or two or three people who trust each other, can now run the operating layer of a real business as software: agents that work every lane on a schedule, one shared memory they all draw from, and a human who decides. Not a lifestyle business, not a side project. A company with a pipeline, books, a research function, and an operations cadence, run by fewer people than a group chat.

The architecture is the org chart

What makes this work is not one clever tool, it is a shape. Five layers, and the order matters. Models at the bottom, because they are interchangeable and should be treated that way. A working memory above them, one queryable source of truth built from your meetings, docs, and records, so every agent and every future teammate draws on the same facts instead of guessing. The agent team above that, each owning a lane a hire used to own. Approval rails above the agents, because nothing customer-facing should leave the building without a human click, and every agent should be graded on a schedule against a written definition of its job. And at the top, you: the judgment, the taste, the relationships. The layers below exist to hand you better decisions, not to take them.

The model bill is a design decision

The second thing that changed is underneath the stack: open-weight models got good. Not good enough to run your company alone, but good enough to carry the high-volume, well-specified work that makes up most of what agents actually do all day: classifying, extracting, summarizing, routing, drafting the first pass of routine things. That work does not need a frontier model, and frontier prices for it add up fast in an agent-heavy company. Hosted open-weight models run that same work for a fraction of the cost, often one to two orders of magnitude less per token.

The mistake is treating this as an either-or. All-frontier is paying premium prices to sort email. All-open-weight is letting your most important customer-facing words be written by the cheapest thing that could write them. The blend is the answer, and the blend has a simple rule: frontier models where judgment carries the business, the analysis, the writing a client will read, anything genuinely hard; open-weight models for the routine volume underneath. Try it in the artifact above. At a realistic agent-heavy volume, moving most of the routine work onto open-weight models cuts the bill by more than half while the judgment work never leaves the frontier.

The rule that keeps the blend honest

One guardrail makes the whole cost conversation safe: a workload only moves to a cheaper model after an evaluation shows the quality held. Not a vibe check, a written test. Define what good looks like for that task, run both models against it, and let the score decide. Do that, and the slider above stops being a gamble and becomes routine cost engineering: every quarter the open-weight models get better, another workload passes the test, and your bill drops without your quality moving. Skip it, and you will not find out what the cheap model broke until a customer does.

This is not theory for us. Foundry runs this exact shape on our own firm: a team of scheduled agents for sales, research, finance, and operations, one universal knowledge layer they all query, evaluation on every agent, and a human approving everything outbound. We are a few-person company that operates like a much larger one, and the architecture above is the reason. If you want this shape mapped onto your own business, at any size, that is exactly what an engagement with us produces.