> ## Documentation Index
> Fetch the complete documentation index at: https://docs.beam.ai/llms.txt
> Use this file to discover all available pages before exploring further.
# Creating Flows
> Build an agent's Flow on the canvas, configure each step from the sidebar, then test and publish it.
A Flow is the set of steps an agent works through when a task starts, from chat or from a trigger. You build it on the agent's **Flow** page, where the steps sit on a canvas and the sidebar on the right holds the settings of the step you select. Every change saves to a draft, and the draft becomes the version the agent runs only when you publish it.
What a Flow is
The Flow page
| Control | What it does |
| - | - |
| **Draft** | Opens **Version history**. It reads **Draft**, or **Active** on a published agent. |
| **Test mode** | Runs the Flow against test connections instead of live systems. |
| **Execute flow** | Sits above the **Trigger** and starts a task on the draft, so you can watch it run. |
| **Publish** | Sits in the banner above the canvas and makes the draft the version the agent runs. |
| **Find a step…** | Searches the canvas from the toolbar at the bottom. |
You can also open the Beam AI panel on this page and ask it to explain the Flow or change a step for you. [Beam AI Copilot](/01-getting-started/beam-ai-copilot/beam-ai-copilot) covers the panel.
Steps and paths
Every Flow starts at the **Trigger** and runs downward along the arrows, one step to the next. A dashed line ending in a **+** marks where the next step goes. Steps come in four categories: **AI** steps follow written instructions, **Integration** steps act in a connected app, **Utilities** are standalone actions, and **Controls** shape the path a task takes: **Loop**, **Condition**, **Wait** and **Exit**.
What a step card shows
| Part | What it tells you |
| - | - |
| Note | The line under the name, saying what the step is for. |
| Amber border | The step is saved but not finished. Hover it to see the missing fields. |
| Warning triangle | An integration the step needs is not connected. |
| Person icon | **Consent required** is on, so the task waits at this step for approval. |
Drafts and versions
A step saves the moment you add it, finished or not, so you can lay out a Flow first and fill in detail later. **Publish** stays greyed out until every step and branch is finished. Each publish adds a version, and you can open an earlier one read-only and restore it as the new draft.
How to use Flows
Changing a Flow needs the **Create & edit** capability, and running it needs **Run**. Without **Create & edit** the Flow opens read-only. Capabilities come from your role, covered in [Roles and Permissions](/05-resources/roles-and-permissions/roles-and-permissions).
Add a step
Every step comes from the same **Add step** window.
Select the **+** at the end of a path, or hover the line between two steps and select the **+** on it.
Choose a category in the left rail, then a step. On **AI**, describe what the step should do and select **Create with AI**, or select **Add blank step** to fill it in yourself.
A new step that still needs settings shows an amber border, and hovering it lists the missing fields. Both can take a moment to appear.
Branch the Flow with a Condition
A **Condition** splits the Flow, and each task takes the branch whose rule it meets.
Open **Add step** where the Flow should split, select **Controls**, then **Condition**.
In the sidebar, set **Condition Type** to **LLM-based** to let an AI model judge, or **Rule-based** for exact matches on earlier values.
Select each **Add condition** label and set when a task takes that branch. To add a branch, select the **+** to the right of the last one.
Each step's settings live in its sidebar.
Select the step on the canvas. To rename it, select the name at the top of the sidebar and type over it.
**Parameters** holds the connection, the **Input variables**, the **Output variables** and, on an AI step, the **Prompt**. **Settings** holds **Consent required** and **On error**, and **Evaluation** holds the scoring metrics.
Select **Configure & test** at the bottom of the sidebar to run the step on sample input and check its output.
Inputs are covered in [Variable Fill](/02-building-agents/agent-configuration/variables-state/variables-state), and scoring in [Evaluation Framework](/04-observability-analytics/evaluation-framework/evaluation-framework).
Choose a step's model
An AI step always has a **Model** section. Other steps show one only when at least one input is filled by AI.
On the **Parameters** tab, open **Model** and select the **LLM**\* field. Models are grouped by provider.
Hover a provider and pick a model from its submenu, then set **Effort** in the same submenu. Higher effort gives the model more room to reason.
Under **Fallback models**, pick the models to try if the first one fails.
**Fallback models** is a Pro feature, locked on a free workspace.
Change a step's icon
A step that uses an integration shows that app's icon until you upload your own.
Select the step, then its icon at the top of the sidebar. Select **Upload**, choose an image of up to 10 MB, and select **Upload** again.
Open the same menu and select **Remove**.
Write a note on a step
A note under a step's name tells you what it is for without opening it.
Select **Add a note to identify this step...** under the step's name, or the note already there.
Press Shift and Enter together for a new line, and Enter to save.
Reconnect a step
Move an arrow to a different step without redrawing it. A branch keeps its condition when you move it.
Hover the arrow and drag it by its arrowhead, the end that points into a step.
Release it over the new step. If that connection is not allowed, the step shakes and the arrow stays put.
Replace or delete a step
Each step's menu swaps or removes it.
Open the three-dot menu on the step card, or at the top of its sidebar.
Select **Replace** to choose a different step in its place, or **Delete** to remove it.
Replacing a **Condition** removes its branches. The steps that were on them stay on the canvas, but nothing points to them until you reconnect them.
Find a step
Search finds a step in the Flow you have open, including a read-only one.
Select the magnifier in the canvas toolbar and type part of a step's name, action or note into **Find a step…**.
Use the arrow keys. The canvas moves to each step and highlights it.
Press Enter to stay on that step, or Escape to go back to where you were.
The **Trigger** holds what starts a task outside chat, such as a timer, a webhook or an app.
Select the **Trigger** step, then **Add trigger** in its sidebar. Pick a trigger under **Built-in**, or an app under **Integrations** and then its action.
Complete its settings, check them with **Configure & test**, and use the switch beside the trigger to turn it on.
Each trigger type is covered in [Triggers](/03-running-operations/task-management/triggers-webhooks/triggers-webhooks).
Test and publish
Try the draft on test connections, then publish it.
Finish every step and branch with an amber border, including the **Trigger**. Hover each one to see what it is waiting on.
Turn on **Test mode** in the header, set the test **Trigger** and **Connections**, and select **Start testing**.
Select **View test runs** in the header to see what the runs produced.
Select **Publish** in the **This flow has not yet been published** banner. If test mode is on, confirming also ends the test session.
Test mode is covered in full in [Test Mode](/03-running-operations/debugging-testing/test-datasets/test-datasets).
Restore an earlier version
Bring back an earlier version as your draft.
Select the version button in the header. It reads **Draft**, or **Active** on a published agent, and **Current draft** appears above **Previous versions** only when you have unpublished changes.
Select a version under **Previous versions**. The canvas shows it read-only.
Select **Restore as draft**. If you have unpublished changes, select **Restore** to confirm they will be replaced.
Troubleshooting
Hover **Publish** to read why. Usually a step, the **Trigger** or a Condition branch is unfinished. If the message names the create & edit permission, ask an admin to publish.
Hover the control to read its placement rule: a **Loop** cannot go inside another loop, a **Wait** cannot go inside a loop, an **Exit** only ends a Condition branch, and two Conditions cannot sit next to each other.
Test mode needs a trigger. If the tooltip asks for one, add it as in [Configure the trigger](#configure-the-trigger). A tooltip about credits means the workspace has run out of them.
The integration the step needs has no connected account. On an Integration step, select the triangle to connect one. On an AI step, connect it from the step's sidebar.
Next steps
Choose how each input variable gets its value.
Define the outputs a step passes on to later steps.
Build the body of a loop and read its output.
Try a Flow against test connections before you publish.
This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.