Who decides
How the gates on a Startup's Decisions shift as people join it.
In every operating model, an Action within its Limits runs, and an Action over them opens an Approval. As Humans join, who answers, who watches, and how far each Agent acts alone all change.
With no humans
An autonomous Startup routes every Approval to the approving Role, whose Agent answers by the Role's Policy. No Action waits on a person.
Two-way decisions run at once. One-way decisions, such as refunds, ad spend, price changes, and outbound campaigns, first wait out a delay window. Its length follows the decision class: how reversible the decision is, and what it puts at stake.
During the window, any watcher can pause, redirect, or cancel the Action. A watcher is another Agent, a Studio partner, or a person's personal agent. When the window closes untouched, the Action proceeds.
Every intervention is a Decision. Pending decisions show in the inbox and the stream as pending items, never as Tasks.
An autonomous store selling standing desks spends $300 a month on Instagram. After a winning test, its Marketing Lead Agent proposes $1,200. The increase is one-way money, so it waits one business day while a Studio partner reads the Evidence and lets it run.
With one human
A single Human holds the Founder Role. Agents still answer most Approvals by Policy, and the Founder answers those above a threshold set per Verb, such as a discount over 15 percent. An Approval left unanswered past its deadline ends its Action canceled, so an absent Founder stops only what needed them.
Autonomy is earned by evals, one decision class at a time. At assist the Agent drafts and the Founder sends. At supervised it acts, and every Action waits in the Founder's pending feed; at full it acts under the normal window.
A class moves up a level when the Role passes 95 percent of its last 50 graded Actions. It drops a level the moment that rate falls below 95 percent.
The two regimes mix in one Startup. A class at full runs behind delay windows as it would with no Human, while a class at assist routes drafts to the Founder. A window is not an Approval: a window proceeds when it closes, and an Approval waits for an answer.
A founder runs a podcast-editing service alone. Her Billing Specialist Agent is at full on refunds up to $100, behind a one-hour window. Her Sales Lead Agent is still supervised on quotes, so each Quote waits in her feed before it goes.
With several humans
Approvals still go to Roles, so each reaches the Human who holds the approving Role. Each Human answers within their own Limits and Mandate, and an Approval above them escalates along the Role's chain to the Founder.
No Worker approves their own Action. An Approval a Human raises routes to the next Role with authority for that decision class. A cofounder's own budget increase reaches the Founder.
Watching multiplies too. Each Human watches the one-way decisions in the Roles they hold, and their personal agent watches on their behalf. Spend states bind everyone, whoever proposed the spend.
A managed bookkeeping firm runs with two cofounders and a staff accountant. The accountant approves refunds above $500 for the Customers she serves, one cofounder approves vendor payments, and the Founder approves anything above both.
What never changes
Limits are Settings that inherit from the Studio down, and a child only tightens them. A guardrail stop never waits for a window or an answer.
Every Approval, intervention, and override lands on the record as a Decision. A Startup is audited the same way in every operating model.
Next
- Who signs: the legal floor and the owner of record.
- Escalation and approvals: stakes tiers, deadlines, and change requests.
- Mandates and limits: where each Limit comes from.