What Does Human in the Loop Mean for an AI Agent?
You keep hearing the phrase. Where the person sits in an AI agent I built live on October 7, 2026.
Human in the loop means a person stays in the work with the AI agent: you check what it produced before it goes anywhere. In the build I taught, the person sits in five places: before anything sends, when the AI agent stops to check, after the first test, on a calendar reminder to read its instructions, and at the connection, where access can be removed.
You've heard the phrase in a meeting, and nobody said where the person actually sits. The build I taught has five. Each one is a setting or a habit you can point at, so you can ask a team to show you theirs.
What do I mean by human in the loop?
I put it this way in a published video interview:
From my published video. One of the things that I am amazed at when people come to my sessions, I teach them about human in the loop. You need a human in the loop with your AI tools, especially agents.
And the reason, with a task that's easy to picture:
From my published video. If we have a standard doc that has the right information, we can then say, "Yeah, autofill and let me check it."
From my published video. But again, it does require a human in the loop.
The AI agent fills in the form, and you check it. In the build I taught, the person sits in five places.
| Where the person sits | What the person does | Page | |---|---|---| | 1. Before anything sends | Gives or restricts permission to send | How to make an AI agent wait for your go-ahead | | 2. When it stops to check | Answers its questions, tells it to continue | What are AI agent guardrails? | | 3. After the first test | Gives feedback with reasons | How to test an AI agent and give feedback | | 4. On a calendar reminder | Reads the instructions it updated | How often to review an AI agent's instructions | | 5. At the connection | Removes or reauthorizes access | How to control what an AI agent can do in Gmail |
1. Before anything sends
Under tool management, I showed the read-only tools first (43:59). Below them are the write and delete tools, and that's where the sending sits:
From the recording, 44:19. I can then give permission or restrict access to sending emails, updating, creating drafts, forwarding emails, et cetera.
The instructions file on the session page also has a rule about sending:
From the file on the session page. Every reply is a draft, saved as a Gmail draft, never sent, until I say the words "send it" in that exact turn.
That's a written rule. The setting that restricts sending is the one in tool management, so ask a team to show you both.
2. When it stops to check
Partway through the build, my AI agent stopped and asked me something:
From the recording, 17:26. If you're following along on my screen, my agent is now asking for input. It's provided multiple choice questions for me.
The session page says it also shows what it found, then stops and checks with you before drafting, and you tell it to continue.
3. After the first test
When I ran the test, I asked the group to answer it with reasons, whatever they thought:
From the recording, 46:26. So if your agent comes back and it gives you a result that you love, tell it why you love the results.
From the recording, 46:47. So don't say try again, actually say, "Hey, you surfaced this and this isn't beneficial and this is why."
4. On a calendar reminder
My AI agent edits its own instructions, so I keep a 15 minute meeting every three weeks to read them (24:51). I said this about it:
From the recording, 25:31. You want to be able to manage and have oversight into what the agent is updating.
From the recording, 25:36. So please put a calendar reminder, at some frequency for you to come back and check that.
5. At the connection
The last place is the plug. In the connectors section, I showed that you can take access away:
From the recording, 45:04. If I click on the three dots, it allows me to switch an account, but it also gives me access to easily remove access.
Access can also stop on its own when a password changes. What Happens to My AI Agent When I Change a Password? covers that.
What does the documentation add?
The builder in my session was Gumloop. On my relationship to Gumloop, I said this:
From the recording, 33:22. I don't work for Gumloop. I found this tool over the summer through Lenny's podcast, and I started using it and have since made myself an ambassador for them. I've created a relationship with Gumloop. I am not an employee.
More on that is in Do you work for Gumloop? (the answer from the session).
Checked on 2026-11-05 against Gumloop's Human in the loop page. It says your approval settings decide which tool calls pause for review, and that a paused AI agent shows an approval card with the proposed action. One detail changes how you set this up: the page lists Always allow as the default for a connector, so a person is in the loop only after you choose a stricter setting. Ask for writes/deletes lets read-only tools run freely and pauses writes and deletes for approval, and the page calls it a great default for most apps. The page also describes choice cards with an Other box, which matches the option I used when my AI agent asked me questions (17:35).
What goes wrong
Instructions that change with nobody reading them. That's why I keep the reminder from place 4, and why I told the group to put one on their calendar (25:36).
A result sent back with no reason. My advice at 46:47: don't say try again, say what it surfaced and why that didn't help.
Where this comes from
This article grows from the session I taught live with Women Applying AI on October 7, 2026, and from my published video on YouTube, whose captions were pulled on October 31, 2026. The quoted passages are my own words, taken verbatim from the recording and the video. The passage marked as from the file on the session page comes from the files published with that session. The steps, tables and checklists around them were drafted by my team's AI agents from those words and my published pages, and checked against the recording before publishing. Screens change, and platform facts that have a public source carry the dated line above.
Build Your First AI Agent in One Hour: the full session with the clips, the prompts and the transcript.
Related reading: How Do You Keep Quality in AI Workflows?, Delegating Execution to an AI Agent Without Delegating Responsibility, What Are AI Agent Guardrails?.
Frequently asked questions
Does human in the loop mean a person approves every step?
Not in the build I taught. The person sits in five places, and only the first is an approval. Gumloop's documentation lists settings from asking each time to asking only for writes and deletes.
Where do I set what an AI agent can send?
Under tool management, where the write and delete tools sit (44:19). The documentation lists Always allow as the default.
How often should the person read its instructions?
I read mine every three weeks, and I told the group daily users may want to check every two weeks (25:10).
Is a written rule the same as a setting?
No. The instructions file on the session page says every reply is a draft, and that's a written rule. The setting that restricts sending is in tool management.
Build something that actually runs your workflow.
A focused, free 60-minute live session with Jacqueline. You build alongside her, on your own real task, and leave with an AI agent that is already running.