Skip to main content
The AI Task node makes a call to a language model (LLM) to process information and return a structured result. Unlike the AI Agent node, which maintains a continuous conversation with the user, the AI Task executes a targeted task and returns its result without interacting directly with the user.

​
When to use AI Task vs AI Agent

​
Configuration

AI Task uses the same configuration fields as the AI Agent:
  • Model: LLM to use
  • Instructions: Prompt that defines what task to execute
  • Temperature: Creativity control (0 = precise, 1 = creative)
  • Knowledge bases: Documents and datastores for context
  • MCPs and Tools: Available external tools
See model and instructions configuration details

​
Hold the conversation while waiting

By default, while the AI Task waits for the model’s response, the user is free: if they write, their message is handled by the project router and can start another flow in the same conversation. A model call can take tens of seconds, so to prevent this open the Advanced tab and enable Hold the conversation while waiting: the user stays in this workflow until the node finishes or the hold expires.
Messages the user sends while the conversation is held do not enter the flow: they are always discarded, so later nodes —including an AI Agent— will not see what they wrote.

​
How long it lasts

The conversation is released as soon as the node finishes, through either the success or the error output, and the flow continues. As a safeguard, the hold lasts at most 5 minutes: if the model takes longer, the user goes back to talking to the router even though the node is still running.

​
When it does not apply

  • Nodes inside a Tool: only the workflow that talks to the user can hold them, so the option is ignored on an AI Task running inside a Tool. In that case, enable it on the workflow’s Tool node; see What is a Tool.
  • Conversations with an operator: if a human agent takes over the conversation, their messages bypass the hold.
  • User already held: if another node —for example, a Pause or a Tool with the same option— is already holding the user, that hold is kept and this node does not replace it.
Tests from the canvas also apply the hold, just like in production.

​
Key difference: structured response

The AI Task does not need the end_function function like the AI Agent. Instead, it directly returns a result in JSON format that you can save in memory to use in subsequent nodes (for example, in a Conditional node). Instructions: The result is saved in memory and you can use a Conditional node to direct the flow based on the detected category.