Agent

The Agent node uses AI to interpret the user’s message, follow your instructions, call tools when needed, and optionally send a reply.

Add the Agent node from the Core section of the node list. It invokes an agent to perform the assigned action on the Workflow path.

Use Agent when the user needs a conversational experience—not only a document search. For example, explain a policy, look up an order, and respond in a consistent tone.

Use Knowledge Base instead when the path should only search stored documents. Agent can still search documents through the optional Knowledge Base Retrieval tool, but only when the model decides a search is needed.

Until you save settings, the canvas shows This node is not configured.

How the Agent Node Works

When the Workflow reaches the Agent node, the model uses the System Prompt, Model, and any configured Tools to process the user’s message. Depending on Enable Answering and Enable Streaming, the user may see a reply from this node.

For example, a shopper asks:

Where is order 1842?

1. The message reaches Agent.
2. The System Prompt instructs the assistant to act as support, look up orders, and never invent tracking numbers.
3. The Model decides it needs data and calls an API Request tool with 1842.
4. The API returns status and tracking.
5. If Enable Answering is on, the shopper sees a reply such as: Order 1842 shipped yesterday. Tracking: 1Z999.
6. If Enable Streaming is also on, that reply appears word by word instead of all at once.

The Agent node has one incoming connection and one outgoing connection. It does not branch by itself. Use If / Else after Agent when the next step depends on a field defined in Output Schema.

Add and Connect Agent

To add an Agent node:

1. Open the Workflow canvas.
2. Find Agent under Core.
3. Add the node to the canvas.
4. Connect the previous node to Agent.
5. Connect Agent to the next node when the path should continue after this step.

A common setup is:

Start → Agent

You can also place Agent after other nodes:

Start → User Feedback → Agent
Start → Knowledge Base → Agent
Start → Agent → If / Else → End

Configure the Agent

Select the Agent node to open its configuration panel.

Select Add or View / Edit to open the Add / Edit Agent Configuration dialog.

Select Save to apply your changes. Select Cancel to close the dialog without applying edits from that session.

Name

The Name is a label for you on the canvas. The model does not see it, and routing does not use it.

The name is required and must contain at least 3 characters.

Examples that work:

Order lookup
Eligibility check

Examples that fail:

AB

System Prompt

The System Prompt is the job brief for the assistant. Describe what the assistant is, what it must do, what it must not do, and when to use each tool by the Tool Name you configure later.

The System Prompt is required.

Example:

You are a customer support assistant for Acme Shop.
Answer only from tool results and the user's message.
If the user asks about an order, call get_order_status with the order number.
If there is no order number, ask for it.
Do not guess tracking numbers. Keep replies under five sentences.

Chat Canvas

Type / in the System Prompt to attach a Chat Canvas you already built (cards, buttons, layout). Attached canvases appear as chips under the prompt. Remove a chip to detach it.

Chat Canvas changes how the reply is presented in chat. It does not replace the System Prompt.

Example: The prompt explains the policy; the canvas shows a product card with image and price.

Provider

Provider selects which company supplies the models in the Model list.

If you switch provider, the Model list updates. If the previously selected model is not available on the new provider, the first available model for that provider is selected.

Example: OpenAI → gpt-4.1-mini. Anthropic → a Claude model from the updated list.

Model

Model is the specific engine for this Agent node. The list comes from your workspace credit catalog, so available names can differ by account.

Each option shows credits / run. If a model is being retired, a deprecation date appears on the option and under the selection.

Model is required.

How to choose:

• Short FAQ-style answers → smaller, lower-credit models
• Long reasoning, many tools, or strict policy → stronger models (higher credits)

The credit value next to the model name is the cost each time this Agent node runs, not a monthly total.

Temperature

Temperature controls how steady versus varied the wording is. It does not change which tools exist; it affects how freely the model phrases the answer.

Default when adding a new Agent: 0.5

Range depends on the model:

Examples:

• Policy, eligibility, or order facts → 0.2–0.4 for consistent wording
• Catchy product suggestions → closer to 0.8–1.0 where the slider allows

If the field is locked or shows N/A, that model does not accept a custom temperature.

Max Tokens

Max Tokens caps reply length (generated text), not the length of the user’s message.

Examples:

• One-sentence status → 300–500
• Multi-step explanation with a short summary → 1000–2000
• Long structured JSON with Output Schema → raise only as far as needed; a high cap allows a long answer if the model produces one

Reasoning Effort

Reasoning Effort appears only for models whose name starts with gpt-5.

It controls extra internal reasoning before the visible answer. Higher effort can improve accuracy on complex decisions but increases latency.

Example: Order lookup with one API → None. Should we refund under this policy given the API result? → Medium or High.

Output Schema (Optional)

Output Schema defines JSON fields the model should return. Skip it when the user only needs a normal chat message.

Use it when a later node must read stable values—especially If / Else.

Each row includes:

Example schema for routing after Agent:

• order_found — BOOL — Whether the API returned an order
• status — STR — Shipping status from the API
• user_message — STR — Short text the user can read

If / Else can then test order_found and send not found versus found down different paths. If Enable Answering is off, Agent can fill this schema without chatting; a later node can present the message.

Enable Memory

Off (default): This run does not include earlier turns from the same chat. Use for a single lookup such as Status of 1842?

On: The model also sees recent messages from this session. Use when follow-up messages depend on earlier context, such as What about the blue one? after the user already named a product.

Context Window Length (1–100)

Shown only when memory is on. Controls how many previous messages to include. Default 10.

Example: Window 6 on a long chat keeps the last few exchanges (including the bot’s lines) without sending the entire history.

Memory applies per chat session, not across different users or conversations.

Tools Configuration

Tools are optional actions the model may call during this Agent node. Without tools, Agent uses the System Prompt, memory (if enabled), and the current message only.

Select Add Tool to add one of four tool types. After a tool exists, edit or delete it in the list. Tools are stored when you Save the Agent configuration.

Tool Name (all types): Use letters, numbers, and underscores only (get_order_status, not get order status). The model invokes the tool by this name. Each tool on the same Agent must have a unique name.

You may add multiple API Request, Knowledge Base Retrieval, and Action Request tools. Shopify: only one integration per Agent.

API Request

An HTTP call the model can trigger.

Worked example: GET https://api.acme.com/orders/{id} to load status.

Send Query Parameters

Adds ?key=value to the URL.

• Static: You set the value (e.g. locale = en-US always).
• Dynamic: The model fills the value from the chat (e.g. order_id from order 1842).

When this option is on, every row needs a name and a value or description. Add at least one complete row.

Send Headers

Same row pattern as query parameters, for HTTP headers.

Static example: Accept = application/json

Dynamic example: A header the model fills from the conversation (uncommon; most headers stay static).

Send Body

For POST, PUT, PATCH, or any call that needs a payload.

Field types: String, Number, Boolean, Object, Array. Each field can be Static (fixed) or Dynamic (filled from the chat; requires a short description).

Example POST body:

• email — String — Dynamic — Customer email from the message
• plan — String — Static — pro

When the switch is on, add at least one field.

Enable Authorization

Adds credentials to the request.

Example: Bearer Token with your API token so get_order_status is accepted by the server.

Leave authorization off only if the endpoint is public.

Knowledge Base Retrieval

The model may search a Curatu Knowledge Base. This is not the Knowledge Base node, which always runs a search on its path.

If the list is empty, create a Knowledge Base under Data first.

Action Request

Use a working curl command from Postman or vendor documentation. Curatu runs it as an HTTP request, not as a terminal command.

Action Command must start with curl (curl, then a space).

Example command:

curl https://api.example.com/tours?q=paris -H "accept: application/json"

Description: Search tours when the user names a destination. Returns titles and dates.

Dynamic Fields: After you paste the command, use Analyze Action Command (or the sparkles control). You get a table with:

• Checkbox — include this field as filled from chat
• Location — query, body, or header
• Name — e.g. q
• Description — e.g. City or destination the user asked about
• Type — string, number, etc.

Make All Dynamic / Make All Static selects or clears every checkbox. Unchecked rows keep the original curl value (e.g. a fixed accept header).

Shopify Integration

One store connection per Agent. A second Shopify tool is disabled (This tool is already included).

Connection Type (new tool only):

• Connect with OAuth — Domain, tool name, actions, then Add Tool. The Agent saves first, then Shopify login runs. After success, the tool appears on this Agent.
• Enter Access Token — Same flow, plus a private Admin Access Token. No redirect.

Store Domain: your-store.myshopify.com

Tool Name example: acme_shopify

Available Actions (select at least one):

Enable only actions this assistant should perform. A storefront FAQ bot may need Search Products and Get Product only—not Cancel Order.

Enable Streaming

Off (default): The full answer appears when generation finishes.

On: Text appears as it is generated.

Streaming affects display only. A node can stream while Enable Answering is off, but there is nothing user-facing to stream in that case.

Example: A long policy explanation often feels faster with streaming enabled.

Enable Answering

Off (default): Agent may call tools and fill Output Schema, but the user does not receive a message from this node. Use when a later node should speak (End, another Agent, User Feedback) or when you only need structured output for If / Else.

On: This node may send the user-facing reply. In many workflows, that is where the path completes for that turn.

Example: Agent sets order_found with answering off → If / Else → End shows We couldn’t find that order.

Configuration Limits

Agent and Other Nodes

Agent at a Glance

Configure the System Prompt and tools first, then set Enable Answering and Enable Streaming based on whether this node should speak to the user and how the reply should appear.