Introduction

wolkvox Studio incorporates the new "Agent in the Loop" component, based on the Human in the Loop (HITL) approach, which allows integrating human agent participation within automated flows built with Agentic Artificial Intelligence.

With this component, a flow can stop at a specific point and send information to a human agent for review and response generation. Once the intervention is completed, the response can be made available to continue with the logic designed in the Routing Point.

This allows building automations where AI and other components execute much of the process autonomously but maintain the ability to involve a person when a situation requires judgment, validation, supervision, or a manual response.

The "Agent in the Loop" component is located in wolkvox Studio, within the "Cognitive" tab, in the "AI Conversational Components" group.

Important: Agent in the Loop is available only for Routing Points of the Agentic Engine type.

Agent in the Loop does not determine by itself what the human should do nor does it automatically execute actions on external platforms. Its function is to send the human agent the information defined in its configuration and receive their response. What happens before and after depends on the other components and the logic built into the flow.

For this reason, it can be combined with components such as MCP, Intent, Agentic Deep Research, Agentic DB, Sentinel, and other wolkvox Studio resources to design autonomous processes that escalate to the human team only when necessary.

What Can Agent in the Loop Be Used For?

The component is especially useful when a process can operate autonomously in most cases, but there are situations that require human participation.

Some examples include:

• Reviewing an AI-generated response before sending it to a customer.
• Escalating sensitive or exceptional requests.
• Validating content before publishing it on an external platform.
• Requesting human authorization before executing a critical action.
• Reviewing cases where an autonomous agent does not have enough information or confidence.
• Allowing a specialist to draft a response for a specific management task.
• Incorporating human supervision in social media, customer service, sales, support, or back-office processes.
• Combining automatic processes with human decisions within the same Routing Point.

Example: YouTube Comment Management

In the shown scenario, a flow is built in Agentic Engine that monitors comments made on YouTube.

Through integration with the MCP component, the flow obtains new comments published on the channel. Subsequently, different components analyze their content and determine what treatment each one should receive.

For example, a user might write: "I really enjoyed learning about the company's culture. Where can I get more information about opportunities to join the team?"

This example uses YouTube, but Agent in the Loop is not limited to this platform. The same principle can be applied to other tools integrated via MCP or information obtained through other wolkvox Studio components.

When the flow reaches Agent in the Loop, an interaction is generated for the configured skill or agent.

In wolkvox Agent, the user must go to the Interactions section located in the left side menu.

The new interaction will appear in the inbox with the subject previously defined in the component. When selected, the agent can review all the information sent from the flow in the right panel.

In this current example, the flow can evaluate the comment and determine whether it can be responded to automatically or if it requires human attention.

When human intervention is required, the route reaches the Agent in the Loop component. This generates an interaction for the agent in charge, including the information configured in the Subject and Body fields.

The agent then receives an interaction with data such as:

• Person who made the comment.
• Comment content.
• Date.
• Video identifier or reference.
• Management status or context.

The content will depend entirely on the variables and data incorporated into the flow.

In the YouTube example, an interaction titled Video Comment Alert is shown, and within it, the comment content and other information collected by the process appear.

To intervene:

1. Open the interaction.
2. Review the information sent by the flow.
3. Click on "Reply."
4. Write the response in the text field.
5. Click on "Send."

The response will be associated with the interaction and can continue within the logic defined in the process.

How to Configure Agent in the Loop

1. Double-click on the component to open its configuration panel.
2. Configure its fields:
   • Subject: Define the title that will allow the agent to quickly identify the purpose of the interaction.
     • For example: Video Comment Alert
   • Body: Define all the information you want to deliver to the agent so they can make a decision or generate a response. This field can include variables from the Routing Point, so it is possible to send information previously obtained through AI components, MCP, data queries, or other stages of the flow.
     • For example, the body could contain:
       • Person's name: $nombre_usuario
       • Comment: $comentario
       • Date: $fecha
       • Video: $video_id
       • Management status: Human attention
   • Transfer to a Skill: Select the skill or agent queue responsible for handling this type of request. All agents enabled for that skill can receive the interaction according to the defined operational rules.
   • Prioritize an Agent ID: If you want to direct the interaction prioritarily to a specific person, enter their agent ID. This is useful when certain decisions need to be reviewed by a specific agent or specialist.
   • Simultaneous Interactions per Agent: Define how many interactions of this type an agent can manage simultaneously. This value allows controlling the workload generated by autonomous processes.
   • The Result Variable (informative only) field displays the predefined variable: $agent_in_the_loop_response
     • This variable cannot be modified.
     • It stores the response sent by the human agent, allowing other components of the Routing Point to use it later.
     • The response accepted by this mechanism is text. Attachments are not considered as a result of the component.
     • For example, an MCP component located later in the flow could receive: $agent_in_the_loop_response
     • and use that information as part of its instructions.
   • After completing the configuration, click on "Save Agent in the Loop."
   • Subsequently, compile the Routing Point.