---
title: Autonomy and approvals
description: How much the Agent may do without asking, who decides that, and what waits for a person whatever the setting.
url: https://salexhub.com/docs/agent/autonomy
type: conceptual
area: agent
section: The Agent
last_updated: 2026-10-09
prerequisites: []
related:
  - https://salexhub.com/docs/get-started
  - https://salexhub.com/docs/developers
  - https://salexhub.com/product/agent
  - https://salexhub.com/product/control
tools:
  - record.update
  - record.create
  - property.update
  - email.send
  - batch.undo
  - role.assign
  - approval.decide
  - subject.erase
plans:
  - free
  - pro
  - business
  - enterprise
checked_against: salexhub/platform be080ea89
---

# Autonomy and approvals

How much the Agent may do without asking, who decides that, and what waits for a person whatever the setting.

> **Availability**: Available on every plan. [Plans](https://salexhub.com/pricing#compare)

The Agent works through the same tools as the rest of SalexHub, as the person it works for: it sees what you can see and can change only what you can change. Autonomy decides one thing, which of its steps still need your click. It never gives the Agent a permission you do not have.

The product says it on the settings page itself: "Every action runs with YOUR permissions either way. This only decides what still needs your click."

## The four modes

| Mode | What the Agent does without asking |
| --- | --- |
| Ask every time (the default) | Reads. "The Agent proposes; nothing happens until you confirm." |
| Auto inside the organization | "Creates and updates records on its own. Messages and email still ask." |
| Auto including replies | "Also answers in messenger threads and sends email from your mailbox." |
| Full autonomy | "Anything you could do yourself, without confirmation", including changes to structure such as objects, properties and pipelines. |

Three things hold in every mode. A step that would leave the organization, such as an email, asks until the mode allows sends. A step the Agent took on its own is still written down: the conversation shows it ran without a click, and the record's activity marks it as done via the Agent. And once the Agent has read text from outside the organization during an answer, such as an email, a form submission or a web page, every change or send it proposes in the rest of that answer is a card, headed "Read before approving".

## Record changes ask property by property

Every property says what AI may do with it. An admin sets it on the property's **AI** tab in **Settings > Data > Properties**, under **What AI may do**: **Never change it**, **Propose a change** or **Change it**. Propose a change is the default. A few built-in properties, such as Do not automate, are Never change it.

When the Agent creates or updates a record, the values it writes are checked against those settings, in every mode:

- a value for a Propose a change property turns the step into a card, even in Full autonomy;
- a value for a Never change it property is refused in every mode, even on a card, and the Agent is told which properties it may not change;
- only when every property it writes is Change it does the step run on its own, if the mode allows it.

So in a new organization, the values the Agent writes when it creates or updates a record arrive as cards whatever the mode. An admin lets it write a property on its own by choosing **Change it** on that property's AI tab. If the Agent itself proposes that change, it is a card in every mode, Full autonomy included.

## What always waits for a person

### Cards in every mode

Changes to who may see and do what never run unattended, Full autonomy included: inviting a member, teams, sharing and restriction rules, field permissions and agent definitions. Neither do workflows, industry packs, connected MCP servers, the organization's memory, a bulk export, the data-protection acts on a person (export and memory review), accepting a suggested value on a record, or anything the public will see, such as a published page, a quote shared by link or the site agent. Each arrives as a card.

### Never the Agent's

Some tools no agent runs in any mode, not even with a card. Among them: erasing a person's data, changing someone's role, deciding or canceling an approval, saving an approval rule, freezing or deactivating a user, creating or revoking API keys, creating webhooks, connecting mailboxes and changing billing. People do those in the app.

### Approval rules hold the Agent too

An approval rule your organization sets, such as invoices above an amount, holds the step for its approvers whoever started it: a person, a workflow or the Agent. The Agent tells you which rule holds the step and who decides.

## Who sets the mode

The mode is set in **Settings > AI > Agent**, under "How much the agent may do on its own", in four layers:

![The four autonomy modes and the rows for each role](/docs/screens/agent/autonomy-light.png "Settings > AI > Agent, under How much the agent may do on its own.")

1. **The organization's mode** is the ceiling. Nothing below it can go further.
2. **A mode per role** sets the default for everyone with that role, and may only be narrower than the organization's mode. A person on a custom role follows the row of the built-in role their membership is based on.
3. **A mode per person** names one person and replaces their role's default, up or down, still under the organization's mode.
4. **The conversation**: the mode chip in the Agent's composer narrows the mode for that conversation only. Someone who may change autonomy can also widen the organization's mode from there, with **Allow for the organization**.

Out of the box only the Owner may change autonomy. It takes two rights: the settings permission, which opens the page, and **Agent autonomy** in the role builder, which Admins do not hold. Both can be given to a custom role. Anyone who opens the page without Agent autonomy sees the setting read-only.

On the same page, **AI for this organization** turns the Agent off everywhere at once: while it is off, nothing calls a model. The switch needs only the settings permission, so out of the box Owners and Admins turn it on and off.

## Approving a card

A card shows what the step will change, with **Approve** and **Decline**. When one answer proposes several changes, they wait together: "2 changes waiting for your approval", with **Approve all 2**.

- A send from your connected mailbox asks first. If your organization turns on an undo window for email (up to 120 seconds, off by default), you can still cancel the send during it.
- In the conversation, each change shows whether you approved it or it ran without a click. In the record's activity, the Agent's changes carry a "via the agent" badge under your name.
- **Undo** takes back the record changes of one answer: what it created is archived, what it edited gets its earlier values, what it archived comes back. It skips any record someone edited since, and it cannot recall an email or a message already sent.

## The limits of one answer

One answer may take up to 12 steps and 5 minutes, within an AI-credit ceiling that depends on the model.

## The tools behind it

The product's registry holds 523 tools, and the class of each one decides from which mode it runs without asking:

| Tools | Class | Runs without asking |
| --- | --- | --- |
| 125 | Reads | In every mode |
| 81 | Writes | From Auto inside the organization, except values for properties set to Propose a change |
| 2 | Sends | From Auto including replies |
| 172 | Structure and steps | Only in Full autonomy |
| 78 | Access, packs and privacy | Never: a card in every mode |
| 65 | People only | Never: no agent runs them |

[Build on SalexHub](https://salexhub.com/docs/developers) shows the same registry from a developer's side.

---

The same page for people: https://salexhub.com/docs/agent/autonomy. Index of this site for agents: https://salexhub.com/llms.txt