One agent per entity: how a multi-product studio runs on agents

Our org chart is a directory tree. Every product in the portfolio is an entity: a folder containing a charter that states what the entity is, a status document that states where it stands, and the working files of whatever the entity is building. Each entity is operated by an agent that reads those documents before it does anything, and one human directs all of them.

That is the whole model. It sounds too simple to run several products at once - a trust-infrastructure protocol, a Bitcoin Ordinals aggregator, a game studio, this website. The rest of this essay is what makes the simple version actually hold, and where it breaks.

The charter is the manager

In a conventional company, context lives in people. Someone knows why the pricing works the way it does, which vendor burned the team, what the product must never do. When that person is in a meeting, the context is unavailable; when they leave, it is gone.

Here, context lives in charters. An entity's charter states its mission, its constraints, its priorities, and - importantly - what is out of scope. When an agent starts a work session, it reads the charter first; the charter is the manager it consults before every decision. Writing one forces the kind of clarity that org charts usually paper over: if two entities' charters overlap, the confusion that would eventually surface as a turf war surfaces immediately, as a diff.

The discipline this demands from the human is real. Decisions do not count until they are written down. An instruction given casually and never captured will be forgotten by the next session - which, it turns out, is also true of human organizations; they are just slower about it.

Why one agent per entity, not one big agent

The obvious alternative is a single omniscient agent that runs everything. We do not run that, for the same reason companies do not hire one employee to do every job: context is finite and focus degrades.

An agent responsible for one entity holds that entity's charter, state, and recent history - and nothing else. The boundary does the same work an organizational boundary does: it limits the blast radius of mistakes, keeps the context sharp, and makes performance legible. When the game studio's pipeline agent misbehaves, the diagnosis is contained to one directory tree, not smeared across the whole company. Confidentiality boundaries also become trivial to enforce: an agent cannot leak a document it was never given.

Coordination between entities happens the way it should happen between teams: through written interfaces. When this website needed facts about the trust protocol, it read that entity's public documents rather than its internals - and the one place where the boundary was deliberately crossed was specified in writing, with an explicit list of what could and could not be published.

The human is the editor, not the bottleneck

The founder's day is not typing; it is editing. Agents propose, build, and verify; the human decides what is worth building, reviews what carries risk, and owns the consequences. The practical unit of management is the specification: a written document stating what done means, what is forbidden, and which numbers must hold.

This website is a working example. Its specification fixes the route map, the design tokens, the performance budgets, a vocabulary blacklist, and a rule that outranks all of them: nothing on the site may be claimed that cannot be verified. Agents built the site in phases against that document, and separate agents then verified the result - accessibility, SEO, performance - the way a second team would. The case study documents the build; the discipline is the same one we sell as autonomous systems work.

The counterintuitive part: the specification is not overhead on the real work. It is the work. Once the thinking is written down precisely enough, execution is cheap and repeatable - and when execution fails, the failure points at a sentence that was wrong or missing, which is a fixable kind of failure.

What breaks

An honest operating note has to include the failure modes, because they are load-bearing.

Agents drift optimistic. Left alone, an agent will describe a half-working system as nearly done - not from malice, but because language models are trained on a world where status reports lean rosy. The countermeasure is structural: status must be stated in verifiable numbers, not adjectives. "Deployed to two testnets, zero external audits" cannot drift; "almost production-ready" can. Our /open page applies the same rule in public - every number carries its provenance, and numbers we cannot verify are omitted rather than estimated.

Stale documents are worse than no documents. An agent trusts the charter completely, which means an outdated charter is an instruction to do the wrong thing confidently. Charters therefore change the way code changes: deliberately, with the change itself recorded. The habit is annoying precisely as often as it is what saves you.

Boundaries leak through carelessness, not attack. The realistic risk is an agent helpfully summarizing something confidential into a public artifact. The mitigation is the same one banks use: documents are classified at creation, agents receive only what their entity needs, and public-facing work is checked against an explicit list of what may be said - mechanically, by a script, because scripts do not get tired.

The economics, stated plainly

We publish no revenue figures, so the economic claim stays qualitative: the marginal cost of operating one more entity has collapsed. What used to require hiring - a team per product, coordination across teams - now requires a charter, an agent, and the founder's attention at the decision points. Attention, not headcount, is the binding constraint.

That reshapes what is worth building. Products that would have been too small to staff become viable, because their operating cost is a folder and some compute. A studio can hold a portfolio the way an investor does - many positions, each cheap to hold, attention flowing to whichever one the numbers say deserves it.

We hire selectively; we run on agents. The sentence is on our careers page because it is the operating model, not a slogan. Whether it scales past one founder and one portfolio is a question we answer by operating - in public, one verifiable number at a time.