Agents ai_agent

AI Agent

Conversational agent with session memory, retrieved-context injection and sub-flow tool-calling.

What does it do?

It's the conversational brain of your flow. An AI assistant that chats with your customer, understands their intent and replies naturally.

How does it work?

Uses advanced language models (LLMs) configured with your prompt (instructions). It reads past conversation memory (session memory) and can use tools (like querying your inventory or searching your Knowledge Base) before replying.

Why use it?

Essential for any support, sales or FAQ bot. Without this node you'd have to program rigid buttons and menus for every possible question.

Inputs

name	type
`text`	string

Outputs

name	type
`content`	string
`model`	string
`turn_index`	number
`streamed`	boolean
`tool_calls`	array

Branches

Fixed branches

success — Success
error — Error

Configuration

model string optional

vision_model string optional

role "copilot" | "chatbot" optional

system_prompt string optional

input_field string required default: "text"

memory_key string required default: "messages"

history_turns integer required default: 20

temperature number required default: 0.7

max_tokens integer optional

reasoning "off" | "low" | "medium" | "high" optional

injectors array<object> optional

stream boolean required default: true

context_memory_keys array<string> required default: []

reply_memory_key string optional

tools array required default: []

max_tool_iterations integer required default: 5

recent_turns integer optional

memory_facts_schema object optional

memory_summary_config object optional

auto_retrieve object optional

Default configuration

{
  "model": "deepseek/deepseek-v4-flash",
  "system_prompt": "You are a helpful assistant. Reply in the user language.",
  "input_field": "text",
  "memory_key": "messages",
  "history_turns": 20,
  "temperature": 0.7,
  "max_tokens": 1024,
  "stream": true,
  "context_memory_keys": [],
  "tools": [],
  "max_tool_iterations": 5
}

Emitted events

node.token
Streamed reply tokens (one event per chunk).
agent.reply
Final reply text with model, length and streamed flag.
agent.tool_call_started
Fires when the LLM requests a tool, before execution.
agent.tool_call_completed
Fires after a tool call resolves (success or error).
tool_call.<name>
Per inline-tool emission. The <name> is the tool's declared name.
agent.tool_iteration_limit
The model hit max_tool_iterations with unresolved calls still pending.