Workflows: building the agent's flow

Workflows is a canvas where you build the path a conversation takes through the agent: which node reads the question, which one decides where it goes, which one answers, and where it ends. It is at Aura AI → Workflows, and it is the only full-bleed screen in the section because a graph cannot work inside a centred column.

Reading the workflow is open to anyone. Saving needs admin or owner.

There is one workflow per agent. There is no list of workflows, no versioning, no separate publish step and no undo. Read the section on saving before you make your first change.

The three parts of the screen

Node Library, on the left, holds the twelve node types you can drag onto the canvas, in three groups.

The canvas, in the middle, with a minimap, zoom controls, an auto-layout menu (Left to Right, Right to Left, Top to Bottom, Bottom to Top) and the Save Workflow button. An empty canvas tells you what to do: drag and drop nodes from the library to get started. An Unsaved changes pill appears the moment the graph differs from what is stored.

Node Configuration, on the right, opens when you click a node. It has a fullscreen toggle for the nodes with a lot in them.

The node types

Raw nodes — the building blocks.

• Agent — an AI agent for processing queries. This is the workhorse: a system prompt, some inputs, an answer.
• Router — routes queries based on conditions. A router has many outgoing routes rather than one next step.
• API Executor — executes an external API call.
• Trained Query — a pre-trained query handler.

Prebuilt nodes — the same machinery, arriving already configured.

• Guardrails — safety and compliance checks. It is a router with two ways out, safe and unsafe.
• Pleasantries — handles greetings and small talk, so your answering node does not have to.
• RAG Agent — retrieval-augmented generation: answers from retrieved documents.
• Response Analyzer — analyses and validates RAG responses. This one comes locked: its inputs, chat history, contextual questions, router settings and LLM settings are fixed, and the panel says so rather than offering controls that would break it.

Deterministic nodes — no model involved, for scripted web-chat steps.

• Message — sends a fixed message and moves on.
• Menu — presents buttons, each routing to a different node.
• Input — collects a single value, with a type of text, email, phone or number, and a required switch.
• Action — the terminal step. Its action type is Create Ticket, Show Link, Collect Lead, or Hand Off to AI.

Hand Off to AI is worth a sentence of its own. It gathers details conversationally rather than through a form, and the guidance you give it is soft: the AI will try to collect what you list and will not block or loop if one is missing. If a value is genuinely mandatory, collect it with an Input node instead.

Configuring a node

Every node has a name, a description and a node type. An AI node adds a preamble — its system prompt — and switches for Use Chat History and Enable Contextual Questions.

Input Variables is a fixed vocabulary rather than free text: Question, Documents, Chat History, Language, Preamble, Brand Voice, User Details, Keyword Glossary and Time. Two of those connect this screen to the rest of the product — Brand Voice and Keyword Glossary are the values you set on the Agent page, so a node that does not take them will not follow your house style however carefully you wrote it.

Save Changes in this panel applies your edits to the canvas only. Nothing is stored until you use Save Workflow on the toolbar.

Edges and routes

Drag from a node's output handle to another node's input handle. A new edge is labelled route when it starts at a router and next otherwise; router edges are drawn dashed.

Double-click an edge label to rename it, which is how you name a router's branches — billing, technical, everything else. Enter commits, Escape cancels.

Deleting a node tidies up after itself: anything that pointed at it is rewired to the end of the conversation rather than left pointing at nothing.

Saving replaces the whole graph

Save Workflow validates first. If anything is wrong, nothing is sent and the errors appear at the top of the canvas, each naming the node and the field — a missing preamble, a menu with no buttons, an action with no action type, a route pointing at a node that does not exist. Fix them and save again.

When it does save, it sends the entire graph. A node that is not on your canvas is deleted upstream. There is no partial save and no merge. Two people editing this screen at once will lose one another's work, so treat it as a single-editor surface and say out loud when you are in it.

On success it tells you how many nodes it saved, then reads the workflow back so what you are looking at is what is stored.