Skip to main content

Human-in-the-Loop

Your AI Agent can do real work on your behalf — send emails, update your CRM, scrape websites, call APIs, run batches across hundreds of rows. The more your AI Agent does, the more you want to stay in control of what it does.

Human-in-the-Loop is how you stay in control. It lets your AI Agent pause mid-conversation, show you what it is about to do (or ask you something), and wait for your decision before continuing.

Think of it as a manager reviewing an employee's work before it goes out. Your AI Agent prepares the action, shows it to you, and waits for a green light.

Two ways your AI Agent can pause​

There are two moments when a human enters the loop:

1. Before a tool runs — approval. You mark a tool as "needs approval". When your AI Agent wants to use that tool, it stops, shows you the exact inputs, and waits. You click Approve and the tool runs, or Dismiss and it does not. Good for anything that writes, updates, sends, or deletes.

2. In the middle of a tool — asking you a question. Some tools are designed to ask. The Ask User Questions tool lets your AI Agent send you a question — single choice, multi-choice, or free text — and wait for your answer before continuing. Good when the AI Agent needs clarification or a decision from you.

Either way, the conversation does not get stuck. Your AI Agent parks itself, waits, and resumes the moment you act.

What a human task actually is​

Every time your AI Agent uses a tool, calls another AI Agent, or runs a batch, Cubeo AI creates a task behind the scenes. Most tasks run and finish in seconds and you never notice them. A task becomes a human task when it pauses and waits for you.

You can resolve a human task:

  • In chat — as an approval or questions card inside the conversation.
  • From the Dashboard — in the Pending Approvals view, which collects every pending task across all your AI Agents.

Where to start​

Common scenarios​

  • CRM writes with HubSpot. Your AI Agent enriches leads and creates contacts — you approve each one before it lands. Walk through it in Your first human-in-the-loop task.
  • Email-sending AI Agents. Your AI Agent drafts outbound emails; you approve the exact text and recipient before it sends.
  • AI Teams with one sensitive subagent. One AI Agent does research (no approval), another updates the CRM (approval on). You only review the CRM actions. See Subagent approvals.
  • Ask-first workflows. Your AI Agent uses the Ask User Questions tool to confirm intent before taking an action. See Answer AI Agent questions.