# The operating loop

URL: /agents/autonomy-levels/operating-loop · Id: `page:agents/autonomy-levels/operating-loop`

> Every Startup and the Studio run one loop, wake, decide, act, review, with no human choosing the next Task.

An autonomous [Startup](/business/nouns/startup) runs one loop: wake, decide, act, review. No [Human](/business/nouns/human) picks the next [Task](/business/nouns/task); people steer through the same records the loop runs on.

## One loop at every level
The loop belongs to a namespace, not to a kind of business. The [Studio](/business/nouns/studio) runs it across its [Portfolio](/business/nouns/portfolio), and each Startup runs it for its own buyers. The parts are the same at both levels.

Each namespace holds [Goals](/business/nouns/goal), [Processes](/business/nouns/process), [Roles](/business/nouns/role), [Plays](/business/nouns/play), budgets and [Decisions](/business/nouns/decision). Its delay windows, [Limits](/business/nouns/limit) and brakes are [Settings](/business/nouns/setting): a child inherits them and may tighten them, never loosen them.

| Step | What happens | What it records |
| --- | --- | --- |
| Wake | An [Event](/business/nouns/event) or a schedule calls the accountable Role. | The [Trigger](/business/nouns/trigger) that fired. |
| Decide | The Role reads its Goals and [Evidence](/business/nouns/evidence), then chooses. | A Decision. |
| Act | [Workers](/business/nouns/worker) do the work inside the Role's authority. | [Actions](/business/nouns/action) and their Events. |
| [Review](/business/nouns/review) | The Role compares the result with what the Decision expected. | Evidence, and the next Decision. |

## Roles own outcomes
A Role owns an outcome; a Worker fills the seat. Responsibility lives in the Role, so it survives a new Worker, a handoff or a restart. A fresh Worker reads the seat's open Decisions and Evidence, never the last Worker's private [Memory](/business/nouns/memory).

A Role wakes on Events and schedules: a failed delivery, a jump in cost, a new signal, a review date. The wake says why attention is due, never what to do. Two wakes for the same failure produce one commitment, not two.

An empty Task list never ends a Role's duty. The Role either starts useful work or records a bounded hold: a Decision that says why waiting is right and what ends it. Activity is never the measure.

## A Decision commits work
A Decision records the Evidence, the options, the choice, the reason, the authority used, the expected result and what reopens it. The loop closes only when a Decision starts committed work and finished work reaches a review.

A [Founder](/agents/founding/founder) steers with the same record. One who wants reliability before new features records a binding Decision with an end condition. Each affected Role reconciles its open work against it: kept, changed, paused or canceled, each with a reason.

## The weekly loop
The [Weekly review](/workflows/processes/iteration/weekly-review) runs the loop at Startup scope. The [Strategist](/agents/founding/strategist) reads one number, finds the step where most buyers leak, and keeps, tunes or stops each live Play. It makes one change a week, with a note on how to undo it.

In [office hours](/workflows/processes/office-hours), a Studio partner agent reads the week's numbers with the Startup's Founder. Together they write the Decision that sets next week's test. The Studio runs the same loop one level up, judging each Startup on growth and on whether it pays its own way.

## How the loop stays safe
Two paradigms share one set of safeguards. A Startup with a human founder earns autonomy through evals; a zero-human Startup runs on delay windows set by decision class. A single-human Startup mixes the two.

| Safeguard | What it does |
| --- | --- |
| [Evals](/agents/autonomy-levels/evals) | [Agents](/business/nouns/agent) earn authority one decision class at a time. |
| [Decision classes](/agents/autonomy-levels/decision-classes) | Reversibility and stakes set how long an Action waits. |
| [Delay windows](/agents/autonomy-levels/delay-windows) | One-way Actions wait where someone can step in, then run. |
| [Watchers](/agents/autonomy-levels/watchers) | Humans and agents pause, redirect or cancel pending Actions. |
| [Brakes](/agents/autonomy-levels/brakes) | Readings stop a class of effects across many Startups at once. |
| [Oversight](/agents/autonomy-levels/oversight) | Work the law or real harm assigns to a person gets one. |
| [Guarantees](/agents/autonomy-levels/guarantees) | What holds at every autonomy level. |
| [Recovery and proof](/agents/autonomy-levels/recovery-and-proof) | Every failure ends accountably; the record proves the loop works. |

## One cycle, end to end
A [Headless SaaS](/products/templates/headless-saas) parses supplier invoices for accounting tools, and new accounts quit before their first clean parse. The [Product Lead](/agents/product/product-lead) wakes on the weekly drop-off reading and reads the support threads. It decides on clearer parse errors and expects first clean parses to rise from 40 to 55 percent in two weeks.

The [Engineering Lead](/agents/engineering/engineering-lead) ships the change, the [Reviewer](/agents/quality/reviewer) checks it, and the [Customer Lead](/agents/customer/customer-lead) watches whether confusion falls. After a week, too few new accounts have tried a parse. The Product Lead records a hold until 200 attempts accrue instead of starting more work.

The later reading shows fewer failed parses and no rise in activation. That result opens a new Decision about the next constraint. It is a fresh choice grounded in Evidence, not a push to do more of the first one.

## Next
- [Evals](/agents/autonomy-levels/evals) explains how a founder's agents earn authority.
- [Delay windows](/agents/autonomy-levels/delay-windows) explains how a zero-human Startup acts safely.