Blog ·
There are two ways to get AI agents wrong. Approve every click, and you've hired an assistant you have to sit next to all day. Approve nothing, and sooner or later it sends, buys or deletes something you didn't mean it to.
“Human in the loop” means deciding where your judgement is actually needed, and building the agent's work around those points. Done well, you spend a few minutes a day on decisions and none on busywork.
Why agents still need a human
- Mistakes compound. A wrong guess on step 3 shapes steps 4 to 40. Catching it early is cheap; catching it at the end isn't.
- Agents can be tricked. A web page or email can carry instructions aimed at the agent. A person notices what doesn't add up.
- Some actions can't be undone. A payment, a sent email or a deleted file stays done.
- Someone is accountable. The agent acts in your name, with your accounts.
- The web is built for people. Sign-ins, two-factor codes and human checks still need a human now and then.
Four levels of supervision
| Level | What you do | Good for |
|---|---|---|
| 1. Watch | Watch it work, live, and step in when needed | The first runs of any new task |
| 2. Approve | It pauses and asks before specific actions | Sending, buying, publishing, deleting |
| 3. Review | It finishes, you check the result before it's used | Drafts, reports, pull requests |
| 4. Spot-check | It runs alone; you look at a sample now and then | Routine jobs with a long clean record |
Different tasks sit at different levels, and a task should only move down the table once it has earned it. An agent can be on spot-check for sorting your inbox and on approve for replying to it.
What should always need your approval
- Spending money, whatever the amount.
- Messages to new people, or anything public: posts, reviews, comments.
- Deleting anything you can't restore.
- Changing security settings, passwords, permissions or who has access.
- Installing software or skills from sources you haven't checked.
- Anything legal or contractual: accepting terms, signing, agreeing to a deal.
How to build approval in
Asking the agent nicely to check with you is a start. Limits it can't talk its way past are better. From weakest to strongest:
- Instructions. “Ask me before sending anything.” Useful, but a confused or tricked agent can ignore them.
- Draft, don't do. Have it produce something for you to act on: a draft email, a pull request instead of a merge, a filled-in cart instead of a purchase.
- Ask in chat. Have it message you on Telegram or WhatsApp with what it wants to do and wait for “yes”. Quick to answer from your phone.
- The harness's own permissions. Coding agents have permission and approval modes that block commands or file edits until you say yes.
- Limits on the account. A read-only role, a spending cap, branch protection, a card with a low limit. These hold even if the agent is fooled, because the agent never had the power in the first place.
Watching an agent work
Logs tell you what the agent says it did. A screen shows you what actually happened: the cookie banner covering the button, the wrong tab, the page that never loaded, the “are you human?” check it has been staring at for ten minutes.
That's why agents that use a real desktop are easier to supervise than ones that work out of sight. Watch the first few runs of any new task from start to finish. You'll learn more about how to word your instructions in ten minutes of watching than in an hour of reading transcripts.
On Cubicles, each agent has its own desktop that you can open from any browser, including on your phone, and watch live while it works. Close the tab and it carries on.
When to take over, and how
Some moments are simply faster for a person:
- Sign-ins and two-factor codes. Sign in once yourself and the session usually stays signed in for the agent.
- Human checks that a site shows before letting anyone in.
- A choice only you can make, like which of two near-identical products is right.
- A loop. If it has tried the same thing three times, it won't work the fourth.
Taking over works best as a clean handoff. Tell the agent to pause, in chat or by stopping the run. Do your part with the mouse and keyboard on its screen. Then tell it what you did and ask it to carry on. Two people moving one mouse at once never ends well, and that includes when one of them is an AI.
Sharing the screen with a teammate or client
Sometimes the right human isn't you. A colleague has the two-factor device, a client wants to see the work being done, or you're demoing the agent to your team. On Cubicles you can share a computer's screen from its page:
- A view-only link that anyone holding it can open, with no account needed. Good for demos.
- View-only for named people: only the emails you invite can watch. Good for clients.
- An invitation to use it: one person, by email, can use the mouse and keyboard with you. Good for a colleague who needs to sign in or fix something.
- An end time for every share: from 30 minutes to 7 days, or no expiry.
For agencies running agents on client accounts, a view-only link per client also builds trust: they can see exactly what the agent does in their store or inbox. See how to run multiple AI agents for keeping clients on separate computers, and how to share your AI agent's screen for every sharing option.
Leave a paper trail
- A work log. Ask the agent to append a line to a log file for every significant action: what, when, and why.
- A summary in chat at the end of every run, even a short one.
- Dated output files instead of overwriting, so you can see what changed.
- Your chat history. Approvals given on Telegram or WhatsApp are a record of who agreed to what.
Moving from watching to trusting
Move a task down a level when it has a clean record, not when you get bored of checking. A simple rule: ten runs in a row with nothing you'd have done differently, then move it one level. If it makes a mistake, move it back up.
Some things never move down: payments, deletions and anything public stay on approve. That's not distrust of the agent. It's the same rule you'd apply to a new hire with your company card.
For the security side of this, see the AI agent security checklist. To try it, create a free Cubicles computer, set up an agent, and watch its first task from start to finish.
Frequently asked questions
What does human in the loop mean for AI agents?
It means a person reviews, approves or can step in at chosen points in the agent's work, such as before it sends, buys or deletes anything, rather than either checking every step or none.
Which actions should an AI agent always ask about?
Spending money, messaging new people or posting publicly, deleting things you can't restore, changing security settings, installing unchecked software, and anything legal or contractual.
Can I watch my AI agent work from my phone?
Yes, if it runs on a desktop you can open in a browser. On Cubicles you can open your agent's computer from any browser, including a phone, and watch it live.
Can I share my agent's screen with someone else?
On Cubicles, yes. Create a view-only link anyone can open, a view-only share for invited emails, or an invitation for one person to use the computer with you, each ending after a time you choose.
How do I stop an AI agent quickly?
Tell it to stop in chat, stop the program on its computer, or stop the computer itself. On Cubicles, stopping a computer ends everything running on it and keeps its files. Revoke its API keys if you think it was compromised.