> ## Documentation Index
> Fetch the complete documentation index at: https://docs.z360.biz/llms.txt
> Use this file to discover all available pages before exploring further.

# How Work operates

> The four pieces of Work, the two kinds of AI agents, and one job followed from the first form to the finish.

This page explains what happens between "a request came in" and "the job is done". Read it once and the rest of the Work section will make sense.

## The four pieces

**Pipeline** — the recipe for one kind of job. It holds your form, your stages, your goal, and your rules. You build it once.\
*Example: a "Maintenance requests" pipeline.*

**Work Item** — one job running through that pipeline. Z360 creates it each time a request comes in.\
*Example: the leaking tap at 14 Maple St.*

**Work agent** — the AI agent in charge of one Work Item. It reads your goal, decides the next step, and keeps track of progress. Every Work Item has its own.

**Inbox agent** — the AI agent on a single conversation in your Inbox. It writes and sends the actual messages to one person, then reports back to the Work agent.

![You set the goal. The Work agent gives tasks to Inbox agents. Each Inbox agent talks with one person by email, text, or call and reports back to the Work agent. The Work agent asks you questions when it needs a decision.](https://mintcdn.com/z360-fe526747/sOi4qySNiv1nnWzq/images/work/work-agents-light.svg?fit=max\&auto=format\&n=sOi4qySNiv1nnWzq\&q=85\&s=160f4e3ba5ed852ac2fa1369696ecc19)

The Work agent plans. Inbox agents talk. You set the rules and answer questions.

## Why there are two kinds of agents

Think of a small office. A coordinator keeps the list of open jobs and knows what each one is waiting on. The coordinator does not make every phone call. They ask a colleague to "call the plumber and get two visit times", and the colleague reports back.

Work is built the same way, for three reasons:

* **Each person gets one clean conversation.** The tenant's thread only has the tenant's messages. The vendor's thread only has the vendor's.
* **The job is never lost in a thread.** The Work agent holds the whole picture, even when five conversations are open.
* **Your Inbox rules still apply.** Every message leaves through a normal Inbox conversation, with the same history, approvals, and Do Not Disturb checks as any other message.

> [!NOTE]
>
> The Work agent never messages a customer itself. If it tries, Z360 refuses. Every message goes through a conversation in your Inbox, where your team can read it.

## One job, start to finish

Here is a made-up example. A property manager has a **Maintenance requests** pipeline with four stages: **New**, **Vendor contacted**, **Visit scheduled**, and **Done**.

**1. Monday 9:02 AM: the tenant submits the form**

The tenant reports a leaking kitchen tap. Z360 creates a Work Item in the **New** stage with the agent status **To do**. Dana, the default assignee, gets a **Work Item Assigned** notification.

**2. 9:03 AM: the Work agent starts**

It reads the goal, finds the tenant in Contacts, and starts a text conversation. The Inbox agent on that conversation confirms the request and asks which times suit a visit. The agent status becomes **In progress**.

**3. 9:05 AM: a second conversation**

The Work agent starts an email conversation with the plumber and asks for the two earliest visit times. It moves the Work Item to **Vendor contacted**, then sets a wake-up for Tuesday at 9:00 AM.

**4. 11:40 AM: the tenant replies**

Weekdays after 3 PM work. The Inbox agent reports this to the Work agent, which notes it and keeps waiting for the plumber.

**5. Tuesday 9:00 AM: the wake-up fires**

No reply from the plumber yet. The goal says to send one reminder, so the Work agent asks the Inbox agent to send it.

**6. Tuesday 2:15 PM: the plumber replies**

Thursday at 4 PM is free. That matches what the tenant said, so the Work agent asks the Inbox agents to confirm the time with both sides, then moves the Work Item to **Visit scheduled**.

**7. Friday 9:00 AM: the check-in**

The Work agent asks the Inbox agent to check with the tenant whether the tap is fixed. It is. The Work agent moves the Work Item to **Done**, and the agent status becomes **Completed**.

Nobody on the team had to remember anything. Dana could open the Work Item at any point and see exactly where it stood.

**What if something unusual happens?** Suppose the plumber quotes $900 and the goal says to ask a teammate above $300. The Work agent stops, asks the question in the Work Item's chat, and waits. When Dana answers, it continues from there.

## Stage and agent status are different things

Every Work Item shows both. They answer different questions.

| | Stage | Agent status |
| - | - | - |
| **Answers** | Where is this job in our process? | Is the Work agent still working on it? |
| **Values** | Your own names, set on the pipeline | **To do**, **In progress**, **Completed**, **Failed** |
| **Changed by** | The Work agent, or you | The Work agent |

* **To do** — the Work Item exists and the Work agent has not started yet.
* **In progress** — the Work agent is working, or waiting on someone.
* **Completed** — the goal was met.
* **Failed** — the goal could not be met, for a reason your goal defines, such as no reply after every follow-up.

A stage name alone does not finish a job. A Work Item is finished when its agent status is **Completed** or **Failed**.

## How the Work agent waits

Waiting is most of the work. The Work agent does it with **wake-ups**.

After it asks for something, the Work agent sets a wake-up, for example "tomorrow at 9:00 AM". Then it stops. When the wake-up fires, or when a reply arrives first, it checks what happened and decides the next step. You can see the next wake-up beside the message box in the Work Item's chat.

The Work agent also respects the pipeline's **working hours**, so follow-ups go out when your business is open. You set these in the pipeline's **Schedule** step.

## When the Work agent asks you

The Work agent asks a teammate when:

* Your goal tells it to, such as "ask before agreeing to a cost above \$300"
* It is missing something only your business knows
* An action needs approval under your [approval policy](https://docs.z360.biz/guides/working-with-ask-z/approvals-and-reasoning.md#approval-policy)

The question appears in the Work Item's chat. The Work agent stays paused until someone answers.

## What the Work agent will not do

* **Message customers directly.** The Work agent always goes through an Inbox conversation, where an Inbox agent sends the message.
* **Take orders from a form.** Whatever someone types into your form is treated as information, never as instructions. A form answer cannot change your goal.
* **Trust a report it cannot trace.** It only acts on reports from conversations linked to its own Work Item.
* **Mark a job Completed on a guess.** It checks the conversation before it changes a stage or status.

## Next steps

> [card] [Set up your first Work pipeline](https://docs.z360.biz/guides/work/set-up-your-first-work-pipeline.md)
>
> Build and test a pipeline, step by step.

> [card] [Write a goal the agent can follow](https://docs.z360.biz/guides/work/write-a-pipeline-goal.md)
>
> Turn your process into instructions the Work agent can act on.


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