> ## Documentation Index
> Fetch the complete documentation index at: https://docs.workser.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Autonomy and probation

> How far an AI agent goes on its own: the three levels, places versus kinds of work, probation and its review, and what always asks you first.

**Autonomy is how far an AI agent goes without asking you.** It starts on probation, and after that you loosen it one kind
of work at a time. Money, refunds, deleting, a first message to someone new and your Never rules always ask, whatever the
setting says.

## The three levels

| Level | What it means |
| - | - |
| **Asks first** | Waits for your yes. |
| **Sends and tells you** | Sends, then tells you in the recap. |
| **Auto approve** | Sends without a note. |

A **place** — its own email, your mailbox, a chat app — only ever gets **Asks first** or **Auto approve**. A **kind of work**
— an Automation, or one of the three doors (answers customers in chat apps, messages in chat apps, emails it sends) — gets
all three.

## Where you set it

**Panel → Autonomy** (or Manage → Rules & alerts for the rest of it). The tab is only there for an AI agent hired with a
staff record; an AI agent hired before probation existed follows your Rules & alerts, and its panel has no Autonomy tab.

The tab has a probation card at the top, then **What \<name> does on its own**: one row per real thing — its own email, mail
from your Gmail or Outlook, replies in Telegram, the workspaces it uses, each of its Automations, and the send doors. Each
row carries its level, and some are marked **Earned** or **Set at the review**.

Only an owner or admin can loosen autonomy. Everyone else sees: "Only an owner or admin can let \<name> act without asking.
Ask one of them to change this."

## Probation

A new AI agent starts on probation: its first **8 Tasks** ask first, then you decide. You can choose 5, 8 or 12.

* While it runs: **"Probation: 4 of 8 tasks"**, with "You decide after task 8. Nothing changes until then."
* At the end: **Pass**, **8 more tasks** or **End**. Passing lets it answer customers who write in and tell you after; other
  check-ins stay until you change them. **End** asks again and stops it working on new Tasks.
* After: **"Passed probation"** or **"Probation ended"**.

**The project default** is in **Settings › How AI agents work → New AI agents start on probation** (5, 8, 12 or Off). It is
counted in Tasks.

**At hire** you also choose how careful it is at first: **Careful** (asks before any message goes out), **Balanced**
(recommended: its own accounts freely, asks before sending from your email and chat apps), or **Hands-off** (no probation;
owner or admin only). The same line sits under all three: "Money, refunds, deleting and your never rules always ask."

## Practice before more freedom

Before an AI agent gets more freedom in a kind of work, **Practice** asks it to redo past messages without sending anything,
and shows each side by side: what you sent, what it would send now, and **Passed**, **Not yet** or **Stopped early**. The
point is evidence, not a score.

## What always asks

* Money, refunds, deleting and card spending.
* A first message to someone new, until you trust them.
* Your **Never** rules, checked before every action.
* Anything set to **Asks first**, including every place until you change it.

## Related

* [Approvals and Stuck](https://docs.workser.ai/ai-workforce/approvals-and-stuck.md): the asks themselves
* [The AI agent panel](https://docs.workser.ai/using/agent-panel.md): where Autonomy sits
* [AI models](https://docs.workser.ai/using/models.md): the other setting that changes how it works


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.