AI Agent - Jelou AI Docs

The AI Agent node is the central component for creating autonomous conversations powered by artificial intelligence in Brain Studio. It allows you to configure an agent that processes user messages, queries knowledge bases, executes tools, and generates contextual responses autonomously. The configuration panel is organized into five tabs:

General — Model and instructions
Tools — Native, custom, and MCP tools
Context — Initial message, conversation history, and external knowledge
Advanced — Save response, fallback model, security, expiration, DLP, and more
Events — Tracking events and evaluation conditions

General

The General tab contains the essential agent configuration.

Model

Select the language model (LLM) the agent will use to generate responses. Consider latency, cost, and task complexity before making your selection. Available models include: OpenAI

GPT 4.1 Mini: a simplified variant of GPT-4.1 optimized for fast responses with lower resource demands.
GPT 4.1: a refined evolution of GPT-4, with better understanding, reasoning, and accuracy.
GPT 4-o (Azure): a GPT 4-o version hosted on Azure, focused on stability and performance in enterprise environments. Supports vision.
GPT 4-o Mini: a faster, lighter version of GPT 4-o, oriented toward speed-first use cases. Supports vision.
GPT 5.2: OpenAI’s latest-generation model with advanced reasoning, larger context, and high precision on complex tasks.

Anthropic (Claude)

Claude 3.5 Sonnet: excellent for complex tasks requiring more elaborate text and extensive contexts.
Claude 4 Sonnet: high reasoning and analysis capability.
Claude 4.6 Sonnet: Claude’s latest generation, with improved reasoning and greater accuracy.

Google (Gemini)

Gemini 2.5 Flash: fast processing with multimodal capabilities. Supports vision.
Gemini 2.5 Pro: advanced reasoning with multimodal capabilities. Supports vision.
Gemini 3 Flash: latest generation of Gemini with optimized multimodal processing. Supports vision.

Meta (Llama)

Llama 4 Scout: agile, low-latency model, ideal for quick ideas and lightweight interactions.
Llama 4 Maverick: high-performance model designed for demanding reasoning and multi-step problem solving.

Models marked with “Supports vision” can process images uploaded in the Context tab. If you need the agent to interpret images as part of its context, select one of these models.

You can also add custom models using the “Add model” button. This allows you to connect your own or third-party models that are not in the predefined list.

Instructions

Define the base behavior of the agent through a system prompt. The instructions determine what role it adopts, what tone it uses, and what steps it follows before responding. Make sure they are concise, concrete, and free of ambiguity. Instructions support variable interpolation using the {{$variable}} syntax, which allows you to dynamically customize behavior based on the conversation context. The field includes a character counter that adjusts based on the selected model, since each model has a different maximum limit. See prompting recommendations and examples

Tools

The Tools tab allows you to add tools that extend the agent’s capabilities beyond text generation. The agent autonomously decides when to invoke each tool based on the conversation context. To add tools, click the ”+ Add tools” button and select the tools you need.

Native tools

These are predefined tools built into the platform. They are marked with the text “(Native)” in the selector:

Tool	Description
Product search	Queries the product catalog configured in the channel
Send interactive message	Renders buttons, lists, and quick replies to the user
Send Call to Action	Sends CTA buttons with URL, with WebView support
Current date and time	Gets the current date and time for a specific timezone
Day of the week	Calculates the day corresponding to a date
Transfer to agent	Transfers the conversation to a human agent or support queue

See detailed documentation on native tools

Custom tools

Tools created by the user from Brain Studio. When adding a custom tool you can configure:

Name and description that the model uses to decide when to invoke it.
Custom input parameters.
Action on execution: behavior after the tool has been executed.

See how to create your first tool

MCP Tools (Model Context Protocol)

Integrate external services using the MCP standard. There are two types of integration:

Native MCP apps

Applications from the Jelou marketplace that are installed directly on the platform. Each MCP app can be configured in two modes:

Integration mode: enables all of the app’s tools at once.
Granular mode: allows you to select individual tools from the app and configure each one separately.

MCP app in granular mode showing individual tools with toggles — MCP app configuration in granular mode with individual tools

Default parameters

In granular mode, each individual tool allows you to configure default parameters. Click on a tool to open its detail view, where you can:

View all tool parameters with their type, description, and whether they are required.
Set default values that the agent will use on every invocation, instead of letting the AI decide.
Use the variable selector to inject dynamic values from the flow (e.g., {{$context.email}}).
Toggle between form view and JSON editor to edit all parameters at once.
Search parameters by name when the tool has many.

MCP tool detail showing parameters with type, description, and default values — Detail view of an MCP tool with its configurable parameters

Each parameter displays an indicator:

Indicator	Meaning
Manual	Has a value set by you — the agent will always use that value
Auto	No value set — the agent decides the value on each invocation

If you leave all parameters as “Auto”, the behavior is identical to before: the agent decides all values based on the conversation context.

External MCP servers

Connections to your own MCP servers via custom URL.

Configure the server URL

Enter the URL of your MCP server endpoint.

Add headers (optional)

Configure custom headers (name and value) for authentication or other metadata.

Select the tools

Once connected, select the available tools you want to enable.

Tool actions

Each tool (native, custom, or MCP) can configure an action that runs after it is invoked:

Action	Behavior
None	The agent continues the conversation normally
End function	The workflow execution ends after using the tool
Pause interaction	The conversation pauses until it is resumed externally

Context

The Context tab centralizes the configuration of the agent’s input context: how each conversation starts, how many previous messages it remembers, and what knowledge sources it consults to generate its responses.

Initial message

Defines the text the agent receives as the first user turn when the node begins executing. This setting determines what information the agent starts with to generate its first response.

Option	Behavior	When to use
Pass the last message	The agent receives the last message in the conversation as the initial input	Most cases: customer service, technical support, general queries
No user message	No user turn is sent to the model; the agent responds based only on its system instructions	When there is no relevant user message to pass to the agent (HSM campaigns, automated workflows, post-data collection)
Custom	Shows a text field with a variable selector to define a custom value	When you need to combine variables or send specific context to the agent

Pass the last message is the default and recommended option for most cases. The agent automatically receives what the user wrote (internally equivalent to {{$message.text}}), enabling a natural conversational experience.

Use case examples

Pass the last message — Customer service A user writes “What are the business hours?” and the agent receives it directly as input, generating a response based on its knowledge base. No user message — When there is no real user query Use this option when the last available message does not represent a real user query. Common scenarios:

Post-HSM campaign: the user responded to a WhatsApp template by tapping a button like “Yes, I’m interested”. That payload is not a query — the agent should start from its instructions and the flow’s context.
Automated workflows: the flow was triggered by a webhook, scheduler, or external API. There is no user message because no user wrote anything.
Post-data collection: previous Input nodes already collected name, order number, etc. The last message is a data point (e.g., "ORD-12345"), not a question. The agent should respond based on the information already stored in memory.

In this mode, the agent generates its first response based solely on its system instructions. If you need the agent to greet the user, configure it in the agent’s instructions.

Custom — Tailored context with variables Allows you to build an input message by combining flow variables. For example, to pass context information to the agent:

Product: {{$context.product_name}}. Query: {{$message.text}}

This is useful when the agent needs additional context beyond the user’s message, such as data collected in previous nodes of the flow.

When selecting Custom, the field supports variable interpolation with the {{$variable}} syntax and has a 100-character limit. Use the variable selector to explore the available variables in your flow.

Remember previous messages

When enabled, the agent includes as context the last N messages of the conversation each time it starts. This allows maintaining conversational coherence when the user resumes a session after an interval.

Enable history

Activate the Remember previous messages toggle.

Define the count

Enter how many previous messages the agent should remember. The valid range is 1 to 50 messages.

Use a low value (5–10 messages) for transactional conversations and a higher value for technical support where the full history may be relevant.

Upload block

The unified upload block accepts documents and images in the same interaction. You can add files in three ways:

Drag and drop files directly onto the block
Click on the block to open the file selector
Paste a URL in the link field at the bottom of the block (supports the same document and image formats listed below)

Type	Supported formats	Maximum size
Documents	`.pdf`, `.xlsx`, `.md`, `.txt`, `.json`, `.csv`	2 MB per file
Images	`.jpg`, `.jpeg`, `.png`	10 MB per file

In addition to .pdf and .csv, you can upload .xlsx, .md, .txt, and .json files as knowledge sources.

Documents are processed one at a time. If you select or drag multiple documents simultaneously, only the first one will be processed and you will receive a notice to upload the remaining ones individually. Images, on the other hand, are processed in batch.

Documents

When you upload a document, a panel opens where you define its name (maximum 30 characters) and an optional description that helps the agent understand the file’s content. Once uploaded, you can edit its metadata or delete it from the list.

Images

Images are uploaded directly without additional steps and appear as thumbnails below the upload block. You can upload up to 3 images per agent.

Click a thumbnail to view it in full screen
Hover over a thumbnail to see the delete button

Images require a model with multimodal capabilities (marked with “Supports vision” in the model list). If the selected model does not support vision, you will see a warning banner and the agent will not process images during the conversation.

Datastores

Connect Datum databases so the agent can query structured information in real time:

Select the datastore to connect
Configure the allowed operations on the database

It is recommended to always load documents in the Context tab to anchor the agent’s responses to your business’s official documentation. Assign clear descriptions to each uploaded file to maximize effectiveness. The model will prioritize information from your files and datastores over its general knowledge, drastically reducing the risk of hallucinations.

External context

Allows resuming this node from an external system via the resume API. Useful for workflows that require external asynchronous processing, such as validations, payments, or approvals.

Enable external context

Activate the Add external context toggle to expand the configuration.

Copy the endpoint

Copy the endpoint URL that your external system will need to call:

POST https://gateway.jelou.ai/workflows/v1/skills/resume

Configure the request

Your external system must send the following request:Headers

{
  "x-api-key": "Your API key",
  "Content-Type": "application/json"
}

Payload

{
  "executionId": "{{$context.executionId}}",
  "message": "string (optional)",
  "pauseInteraction": false
}

The executionId is available as {{$context.executionId}} within the flow and is valid for 24 hours.

Memory management

The Memory management section groups a set of sources that, when enabled, connect native tools so the agent can save and recall information, read and update the contact’s CRM data, share variables between nodes, and access platform data (user, channel, and company) as context. Each source is an independent toggle: you enable only the capabilities the agent needs. The memory and CRM sources expose operations the agent invokes autonomously, while the platform context blocks are injected automatically into the prompt.

Source	Tool	Type	What it does
Internal memory	`memory_manager`	Operations	Lets the agent save, recall, and delete information from the conversation to reuse it later.
Inbox CRM	`crm_connect`	Operations	Reads, sets, and updates the contact’s CRM data (name, phone, tags, notes, and more).
$context	`flow_context`	Operations	Reads and writes context variables to share data between nodes.
User information	`platform_reader`	Context block (read-only)	Read-only access to the current user’s data: name, phone, attributes, and metadata.
Channel information	`platform_reader`	Context block (read-only)	Read-only access to the channel’s data: configuration, identifiers, and public attributes of the connected channel.
Company information	`platform_reader`	Context block (read-only)	Read-only access to the company’s data: name, plan, and public attributes injected as prompt context.

Configurable operations — Each operation of the operations-type sources (Internal memory, Inbox CRM, $context) has a popover where you define how its parameters are resolved in AI, variable, or static mode, along with a post-execution behavior (Continue, End, or Pause). CRM permissions — The Inbox CRM source manages permissions per group (contact data / tags) with separate read and write access. You can restrict which tags the agent may access. Memory space — The Internal memory source includes an optional Memory space field to isolate or group saved entries, so different contexts don’t mix. Read-only context blocks — The User information, Channel information, and Company information sources are read-only. They are injected as context blocks into the prompt and do not consume a tool call. The descriptions above are illustrative; see the native tools reference for the exact fields each block exposes.

Internal memory is per-user working memory: it recalls the same user’s data across conversations (working memory, with limited retention). It is not a permanent store.

Inbox CRM requires at least one connected channel for contacts to exist to read from or write to.

See detailed documentation for these tools

Advanced

The Advanced tab contains additional settings for finer control over agent behavior.

Save response

Allows you to store the agent’s last response in a variable for use in subsequent nodes of the flow.

Enable save response

Activate the corresponding toggle.

Define the variable name

Enter the name of the variable where the response will be stored (for example: ai_agent_response).

The variable will be available as {{$context.ai_agent_response}} in subsequent nodes of the flow.

Fallback model

Select an alternative model that will be used automatically if the primary model fails or is unavailable.

The fallback model cannot be the same as the primary model.

Processing options

Independent options to enable additional capabilities for messages the user sends during the conversation:

Option	Description
PDF support	Allows the agent to read and process PDF documents sent by the user during the conversation (text only, not images inside the PDF)
Read image URLs	Enables the agent to process images that the user sends as a URL with descriptive text during the conversation
Quick Reply Payload support	Allows the agent to interpret the payload of quick replies from interactive messages

These options apply to files that the end user sends during the conversation. Images and documents you configure as agent context are managed from the Context tab.

Multi-message response

Splits the agent’s response into separate bubbles when it contains paragraph breaks (\n\n), media URLs, or lists. Image, video, audio, or document URLs are sent as independent media bubbles according to the channel, while numbered lists are kept together.

Enabled by default for agents created after April 24, 2026. Earlier agents keep the previous behavior (single bubble) until you manually enable the toggle.

Security (Guardrails)

Configure the agent’s level of protection against misuse, prompt injection, and out-of-scope requests.

Enable security

Activate the security toggle to access the configuration options.

Select security level

Choose the preset that best fits your use case:

Level	Description
Low	Minimum protection, greater response flexibility
Medium	Balance between security and flexibility (recommended)
High	Strict protection, restricts out-of-scope responses
Critical	Maximum protection, ideal for regulated contexts

Model Armor (optional)

Enable an additional protection layer that filters model inputs and outputs to detect malicious content.

See complete security and guardrails guide

Expiration

Configure a time limit for the agent session. If the user does not respond within the configured time, the session expires automatically.

Enable expiration

Activate the expiration toggle.

Configure duration

Define the time and select the unit:

Minutes: range from 1 to 1200
Hours: range from 1 to 20
Default value: 8 hours (28,800 seconds)

DLP (Data Loss Prevention)

Enable automatic detection and masking of sensitive information in agent conversations.

Enable DLP

Activate the DLP toggle to access the configuration.

Select sensitive data types

Choose which types of information should be automatically detected and masked:

Data type	Default replacement value
Credit card number	`[CreditCardNumber]`
Credit card track number	`[CreditCardTrackNumber]`
Email address	`[EmailAddress]`
Financial account number	`[FinancialAccountNumber]`
IP address	`[IpAddress]`
Location	`[Location]`
Geographic coordinates	`[LocationCoordinates]`
Phone number	`[PhoneNumber]`
Date	`[Date]`
Person name	`[PersonName]`

Customize replacements (optional)

Modify the replacement value for each data type according to your needs.

The Credit card number type is always selected by default and cannot be disabled.

Follow-up message

Lets the agent proactively resume an inactive conversation by sending a reminder to a user who stopped replying. When you turn on the Enable follow-up message toggle, a settings button appears that opens a modal where you define how the reminder behaves.

Enable follow-up message

Turn on the toggle to reveal the settings button. Click the gear icon to open the Follow-up message modal.

Choose the mode

Select how the reminder is generated:

Automatic: the agent writes the reminder on its own based on the conversation.
Manual: you write an instruction that defines how the agent should behave when resuming. The field accepts up to 900 characters.

Set the limits

Adjust the two controls that prevent the agent from over-nudging:

Field	Description	Values
Wait between wakeups	Idle time before sending each reminder.	5, 10, 15, 20, 25, 30, 45, 60 or 120 minutes (default 5)
Stop after # attempts without reply	Maximum number of consecutive reminders without a reply before stopping.	1 to 5 (default 2)

The available options are capped by the agent’s expiration time: you can only pick values lower than that time. Once the session expires, the agent stops sending reminders.

Events

The Events tab groups the node’s instrumentation into two sections: Event configuration and Evaluation conditions. Both feed your metrics and performance dashboards.

Event configuration

Registers tracking events that fire every time the node runs, with a name and up to 5 custom properties per event. This section works the same as in the other instrumentable nodes. See the full node events guide

Evaluation conditions

Evaluation conditions let you measure your agent’s quality and outcomes. You define through a prompt what must be true in a conversation; when the agent ends the conversation, it analyzes the full history against each active condition. For every condition that is met, an event named after the condition fires, ready to analyze in Metrics.

Add a condition

Click “Add condition” and fill in the two fields:

Field	Description	Limit
Name	Identifies the condition and defines the name of the event that fires	60 characters
Instruction (prompt)	Describes what must happen in the conversation for the condition to be met	900 characters

Enable or disable conditions

Each condition has an independent toggle. Disabled conditions are neither evaluated nor fire events, without needing to delete them.

You can create up to 10 conditions per agent, search them by name, and edit or delete them from each item’s menu.

Write each condition for a concrete, verifiable scenario, for example: “the user confirmed an appointment with a date and time”. To distinguish multiple scenarios, create separate conditions — each one fires its own event.

The evaluation runs only when the agent ends the conversation (for example, through a tool with the End function action). If the session expires due to inactivity, the evaluation does not run. Test conversations from the tester do not generate evaluation events either.

Use cases

Measure booked appointments in a sales agent

A workflow with an agent that books medical appointments: the appointment_booked condition is configured with the instruction “It is met if the user confirmed an appointment with a specific date and time”. Every conversation that ends with a confirmed appointment fires the appointment_booked event, and the metrics show how many appointments the agent generates per day.

Detect unsatisfied customers

A customer service workflow: the unsatisfied_customer condition is configured with the instruction “It is met if the user expressed annoyance or frustration, or asked to talk to a human without being attended”. The event makes it possible to quantify problematic conversations and prioritize improvements to the agent’s instructions.

Configuration example

A typical configuration flow for a customer service agent:

Configure General

Select GPT-4.1 as the model and write clear instructions defining the agent’s role, tone, and scope.

Add Tools

Enable Transfer to agent to escalate complex conversations, Product search if the agent handles catalog queries, and Send interactive message to show options to the user.

Configure Context

Select Pass the last message as the initial message, upload the FAQ document in PDF format, add reference images if the model supports vision, and connect the product datastore if applicable.

Adjust Advanced

Configure a fallback model, set expiration to 30 minutes, enable security at medium level, and activate DLP if financial or personal data is handled.

​General

​Model

​Instructions

​Tools

​Native tools

​Custom tools

​MCP Tools (Model Context Protocol)

​Native MCP apps

​Default parameters

​External MCP servers

​Tool actions

​Context

​Initial message

​Use case examples

​Remember previous messages

​Upload block

​Documents

​Images

​Datastores

​External context

​Memory management

​Advanced

​Save response

​Fallback model

​Processing options

​Multi-message response

​Security (Guardrails)

​Expiration

​DLP (Data Loss Prevention)

​Follow-up message

​Events

​Event configuration

​Evaluation conditions

​Use cases

​Configuration example

General

Model

Instructions

Tools

Native tools

Custom tools

MCP Tools (Model Context Protocol)

Native MCP apps

Default parameters

External MCP servers

Tool actions

Context

Initial message

Use case examples

Remember previous messages

Upload block

Documents

Images

Datastores

External context

Memory management

Advanced

Save response

Fallback model

Processing options

Multi-message response

Security (Guardrails)

Expiration

DLP (Data Loss Prevention)

Follow-up message

Events

Event configuration

Evaluation conditions

Use cases

Configuration example