|
Short on time, or new to AI? We can tailor this article to you.

You hand a new colleague a task on Monday. By Monday afternoon they have rewritten a page, emailed a supplier and moved a folder. Some of that is great. Some of it you wanted to see first. Nobody told them where the line was, so they guessed.

An agent has the same problem, with one difference: it works faster than any colleague, and it never gets tired of guessing. So the line has to be drawn on purpose.

Here is where we draw it for human in the loop AI agents: an agent should ask before it does anything you cannot undo, anything the public will see, or anything that spends your money. Everything else, it should just do.

Why asking for everything fails

This is the same question as agent or chat: how much of the work you hand over, and how much you keep.

The first instinct is to make the agent ask about every step. Give people ten cards in a row and they stop reading and press yes. An approval you press without reading is worse than no approval, because it feels like a safeguard and is not one.

The opposite instinct is to let the agent run free. That works until the one time it does not. A draft that is wrong costs you a minute. A post that is live, a record that is deleted or a campaign that is sent to a list costs you an afternoon, or a customer.

This is what people call keeping a human in the loop: the agent does the work, and a person decides the few things that matter. So the useful question is not "should my agent ask?" It is "which actions deserve a question?"

A rule of thumb you can use today

Before you let an agent act on its own, ask three things about the action:

  1. Can I undo it? Editing a draft, yes. Deleting a project, no.
  2. Who will see it? A note in your workspace stays with you. A published page, a social post or a public link does not.
  3. Does it cost money? Reading a file does not. Building an installable app does.

If the answer to any of the three is the wrong one, the agent asks. If all three are fine, it works, and you review the result afterwards.

What human in the loop looks like in Crewdle

In Crewdle the agent shows a card in the conversation. The card says in plain words what it is about to do, and it shows the details: the exact changes to a website's database, the exact code in a script, the price of a build. You choose Approve or Deny.

A few examples of what waits for you:

  • Going live. Publishing a website or app, rolling it back, or adding a custom domain. Also publishing an AI automation or a Roblox game, and posting on Facebook or Instagram.
  • Deleting things. Deleting or moving a website project, or deleting in Xero or Shopify. For several of these, only a person can approve, never another agent.
  • Code. With Google Apps Script, you approve the exact code before it is changed, and you approve a new version before it is published. See the Google Apps Script page.
  • Your customer lists. With Klaviyo, every change an agent makes needs your approval, and agents have no way to send a campaign or delete anything. See the Klaviyo page.
  • Your files. In OneDrive, SharePoint or Google Drive, files the agent created itself are free to change. For any other file, there is one card per file.
  • First contact. The first email an agent sends from its own address to someone new waits for you.

Notice what is not on the list. The agent drafts, researches, reads, sorts and plans without asking.

Always approve, and where it stops

For routine actions, you can press Always approve. The agent then does that kind of action without asking, for that agent only. Saving a memory note is a good example: low stakes, easy to review later.

For what is public, costs money or cannot be undone, there is no Always approve. Every action is its own decision. A card in front of you at that moment beats a setting you forgot. The day you have forgotten what you agreed to is the day you want a card in front of you.

What happens when you say no, or say nothing

If you press Deny, the action does not happen. The agent is told, and it either carries on with something else or asks you how to proceed.

If you do not answer, the action is refused; silence never counts as a yes. The agent then has to ask again with a new card.

You are not the only safeguard

A card is one layer. Each agent can have a monthly budget, and at the limit it pauses and asks you whether to raise it. Admins can set budgets by role. You can stop a task from Workspace, then Tasks, or pause a whole project. Every tool an agent uses is recorded.

In a project with a manager, the manager proposes a plan first, and nothing starts until you press Approve and start. Only the manager asks you questions, so you are not buried in cards from five agents. How approvals and tasks work.

The stance, in one line

An agent that never asks is a risk. An agent that always asks is a chore. The good version asks rarely, about the right things, and shows you exactly what it wants to do.

If you want to see it, add an agent, give it a job that touches something real, and watch which cards show up. If one feels unnecessary, or one is missing, we want to know.