---
title: Edges and conditions
description: "How an edge decides when the conversation moves on: the four transition types, forward and backward paths, and how to write a condition."
---

<Badge variant="accent">Flow agents</Badge>

An edge joins two nodes and carries the condition under which the conversation
follows it. Choosing the label on an edge opens **Configure Edge Connection** in
the panel on the right. An edge with no condition is always followed; an edge with
a condition is followed only when the condition is met, and a node with several
outgoing edges is where the flow branches.

:::tip[Write conditions as a person would say them]
The most common transition type is a natural-language condition. Describe the
situation, not the words: `The caller has confirmed their date of birth` works better
than a list of phrases to match.
:::

![The edge editor for the edge between a Prompt node and End Call.](/media/build/edge-panel.webp)

## How it works

While the conversation is at a node, the agent keeps checking the node's outgoing
edges. As soon as one edge's condition holds, the conversation moves to that edge's
target node and the target's instructions take over. Natural-language conditions
are judged by a model from the transcript so far; tool-result and variable
conditions are checked against facts the flow already has.

An edge is created when you add a node from a **+** handle. Drag from a handle to
another node to connect two existing nodes.

## Configuring an edge

The panel has two tabs, **Forward Path** and **Backward Path**, with the same
fields. The forward path is the normal direction, from the node you started at to
the node you added. The backward path applies when the flow returns along the edge,
for retry or correction loops; it defaults to **None (Always)**.

| Field | Type | Meaning |
|---|---|---|
| **Transition Type** | Select | One of the 4 types below. |
| **Edge Label** | Text | The text shown on the canvas, for example `User confirms`. Shown for natural-language conditions. |
| **Natural Language Condition** | Text | The condition itself, with **Natural Language Condition**. |
| **Tool Execution Result** | Switch | With **Tool Result Condition**: on, it reads **Success** and the path is taken when the tool succeeded; off, it reads **Failure** and the path is taken when it failed. |

A new edge's transition type depends on the node it leaves: **Natural Language
Condition** from a **Start Call** or **Prompt** node, **Tool Result Condition**
with **Success** from a **Webhook** node.

**Save Changes** applies the edge; **Delete Edge** removes it after a confirmation;
**Cancel** discards the panel. The flow is written when you choose **Save** at the
top of the page.

## Transition types

![The 4 transition types.](/media/build/edge-transition-types.webp)

### None (Always)

The edge is always followed without any conditions. Use it when a node has a single
next step: after a **Webhook** node that must always be followed by a confirmation,
or from the last **Prompt** node to **End Call**. If a node has one always-edge and
other conditional edges, the always-edge wins as soon as the node's instructions are
done, so keep it for nodes with one way out.

### Natural Language Condition

The default for edges leaving a **Start Call** or **Prompt** node. The agent
judges the condition
against the conversation. Two fields: **Edge Label**, which is what the canvas
shows, and **Natural Language Condition**, the sentence that is judged. The panel's
own examples: `When the user wants to talk to support`, `If the customer mentions a
billing issue`.

Write one condition per edge, make sibling conditions mutually exclusive, and give
the node a fallback edge (`The caller does not want any of these`) so the
conversation never stalls.

### Tool Result Condition

For edges leaving a **Webhook** node, or a **Prompt** node that calls tools. The
edge is taken when the tool call succeeded (**Success**) or failed (**Failure**).
Pair two edges, one of each, so a failed lookup goes to a node that apologises
and takes a message instead of stopping the flow.

### Variable Condition

Branches on the value of a variable extracted on a **Prompt** node's **Extract**
tab. **Select Variable** lists the variables extracted on the node the edge leaves
and on the nodes before it, each with its type and the node it comes from; with
none there is nothing to pick, so define the variable first. See
[Nodes](/build/flow-agents/nodes#extract).

| Field | Meaning |
|---|---|
| **Select Variable** | The variable to test. |
| **Operator** | Depends on the variable's type. Text: **Equals**, **Not equals**, **Contains**, **Does not contain**, **Starts with**, **Ends with**, **Is one of**, **Is not one of**, **Is empty**, **Is not empty**. Numbers: **Equals**, **Not equals**, **Greater than**, **Less than**, **Greater than or equal**, **Less than or equal**. Boolean: **Equals**, **Not equals**. |
| **Compare Value** | The value to compare with: a **True**/**False** switch for a boolean, a list of the variable's possible values when it has them, comma-separated values for **Is one of** and **Is not one of**, or a free value. Not shown for **Is empty** and **Is not empty**. |

**Condition Preview** shows the condition as written, and the edge's label is set
from it, for example `{{order_status}} equals shipped`.

## Behaviour and limits

- Conditions are evaluated at the node they leave. An edge cannot see what happens
  after its target.
- A natural-language condition is a judgement, not a match. Test it in **Chat** with
  a few phrasings of the same intent before publishing.
- Edge labels are for people. The agent reads the condition, not the label, so an
  edge labelled `Yes` with an empty condition is not a condition.
- **Auto Layout** on the toolbar redraws the canvas along the edges when branches
  make it hard to read.

## Related

- [Flow editor](/build/flow-agents/flow-editor)
- [Nodes](/build/flow-agents/nodes)
- [Dynamic variables](/build/dynamic-variables)
- [Testing in the browser](/test-and-improve/testing-in-the-browser)