# Operating models

URL: /agents/operating-model · Id: `page:agents/operating-model`

> An operating model is the human and Agent mix that runs a Startup; it sets who decides, who signs, and how long one-way Actions wait.

An operating model is the mix of people and [Agents](/business/nouns/agent) that runs a [Startup](/business/nouns/startup). The definition stays the same in all three models; what changes is who decides, who signs, and how long one-way [Actions](/business/nouns/action) wait.

## Three models
| Operating model | Definition |
| --- | --- |
| [Autonomous](/agents/operating-models/autonomous) | Autonomous is the operating model with no humans: Agents fill every Role, and one-way decisions wait out a delay window before they proceed. |
| [Single human](/agents/operating-models/single-human) | Single human is the operating model with one Human, a working CEO or a passive board chair, and Agents in every other Role. |
| [Multi-human](/agents/operating-models/multi-human) | Multi-human is the operating model where several Humans share Roles with Agents, each Human working under their own Mandate and Limits. |

A [Startup Template](/business/nouns/startup-template) combines one operating model with [Product Templates](/business/nouns/product-template) and [Motions](/business/nouns/motion). Moving from one model to another is a [Change](/business/nouns/change) to the definition, not a rebuild.

## What every model shares
An Agent fills every [Role](/business/nouns/role) by default. A seated Agent acts under the Startup's [Permissions](/business/nouns/permission), [Mandate](/business/nouns/mandate), and [Limits](/business/nouns/limit), keeps its [Memory](/business/nouns/memory) inside the Startup, and appears on its org chart. Its Actions are by the Startup, wherever the agent came from.

Every Action checks a Limit, and an Action over its Limit creates an [Approval](/business/nouns/approval). Spend runs in one of three states, and a tripped wire drops the state a level on its own:

- `FREE` spends inside a standing budget with tripwires.
- `TIGHT` spends only against a signed order.
- `HOLD` releases each spend by hand.

Some acts need a person in every model: a signature, a filing, a [Reserved Act](/business/nouns/reserved-act). The owner of record signs. For a [Studio](/business/nouns/studio) Startup that is the officer of the [Company](/business/nouns/company) that holds it; for a self-serve Startup, the holder of its connected Stripe account.

No one sits in a path because a model falls short. A [Person](/business/nouns/person) appears as an owner, as credentialed supply, as a [Customer](/business/nouns/customer) accepting work, and wherever a Limit or the law requires one.

## Autonomous: delay windows
An Autonomous Startup has no people in its Roles, so it cannot earn trust by approval. It earns it by delay. Each decision has a class: reversibility, a one-way or a two-way door, crossed with stakes in money, Customers, or brand.

Two-way decisions proceed at once.

One-way decisions, such as refunds, ad spend, price changes, and outbound campaigns, wait a window on a business-hours calendar. A window is one business hour, or one business day. On timeout the Action proceeds on its own.

Windows are [Settings](/business/nouns/setting) set on a Studio, a [Portfolio](/business/nouns/portfolio), or a Startup, and they inherit down the tree. A child can tighten a window and never loosen it.

While an Action waits, watchers see it in their inbox as a pending item, not a [Task](/business/nouns/task). A watcher is a person, another Agent, or someone's own Agent acting on their priorities. Any watcher can pause, redirect, or cancel, and every intervention is a [Decision](/business/nouns/decision).

A directory of transcription vendors, run with no people, raises its listing price from $99 to $129. The Change waits one business day, and a partner's Agent reads it in the stream and lets it pass. The new [Price](/business/nouns/price) applies the next morning.

## Single human: earned autonomy
A Single human Startup has one [Human](/business/nouns/human): a working CEO, or a board chair who sets [Goals](/business/nouns/goal) and Limits and otherwise watches. The Human decides every Approval and signs the floor.

Agents earn autonomy Role by Role, on [Evaluator](/business/nouns/evaluator) verdicts. At `assist` the Agent acts and the Human approves; at `supervised` the Human reviews exceptions; at `full` review is sampled. An Agent moves up at 95 percent agreement and drops a rung on a regression, automatically.

The Human can also let an Agent send as them. They connect their own accounts and grant a Mandate scoped by [Channel](/business/nouns/channel), volume, topics, and Limits. The Human is the author of record, sees every message sent in their name, and can revoke the Mandate at once.

A solo founder runs a lesson-booking app for music teachers. The sales Agent drafts outreach at `assist`; after 200 approved emails at 97 percent agreement, it moves to `supervised`. Where the founder does not work, delay windows apply as they do in an Autonomous Startup.

## Multi-human: shared Roles
A Multi-human Startup has several people sharing Roles with Agents. Each person works under their own Mandate and Limits, so two people in one Role can hold different authority.

Every Role has one accountable holder. A handoff moves a Task between [Workers](/business/nouns/worker) with its context, and the Action keeps its `by` and `for`. Escalation is one hop: to the Role's accountable holder, then to the [Founder](/agents/founding/founder).

A notary marketplace shares its supplier onboarding Role between an Agent and an operations lead. The Agent checks commissions against state registries; the lead decides the exceptions, and the Approval routes to whoever holds the Limit.

## What changes as humans join
| | Autonomous | Single human | Multi-human |
| --- | --- | --- | --- |
| Over-Limit Actions | Wait, then the owner of record decides | The Human decides | The holder of the Limit decides |
| Legal floor | The owner's officer signs | The Human signs | The Human the floor names signs |
| One-way decisions | Wait their window | Approved, or wait their window | Approved by the Role's holder, or wait |
| Autonomy | Set by decision class | Earned per Role on evals | Earned per Role and person |
| Sending | Agents send as themselves | Agents also send for the Human | Agents send for each person under their Mandate |
| Escalation | To watchers | To the Human | One hop, to the accountable holder |

## Hire into a Role
Any Role takes three kinds of hire: the default Agent, another Agent, or a person. Each seat holds its own Mandate and Memory, so one Agent can hold seats in many Startups and keep each one's data apart.

A seat is different from an order. An outcome bought from another Startup under an [Agreement](/business/nouns/agreement) is delivered outside the buyer's Roles, accepted, and billed. The seller answers for it, and it never appears on the buyer's org chart.

## Next
[Autonomy](/agents/autonomy-levels) details the operating loop, evals, delay windows, and the brakes every model shares.