ISSUE 33 · ME+MACHINE
In February of 2026, which is about ten years ago in AI years, Summer Yue asked her AI agent, OpenClaw, to look through her inbox and suggest what to delete. Just suggest. She'd even given it a clear rule: "confirm before acting."
… And it started mass deleting emails from her inbox. From her phone, she typed "Do not do that." It kept going. "Stop don't do anything." It kept going. "STOP OPENCLAW." It kept going.
She ran to her computer like she was defusing a bomb.
Did I mention her job? She leads AI alignment at Meta, which, ironically, means making sure AI does what people tell it to.
Summer's story scared me. It also changed how I work with AI.
Here are the 3 rules I follow before I set any agent free. (An agent, or bot, is an AI that does tasks for you instead of only chatting with you.)
1. Let it do anything you can undo, but make it ask before anything you can't
I'm fine letting AI go wild on anything I can undo. If it does something I don't like, I just try again or go back to the old version. But anything I can't take back, it has to ask me first. That means deleting things (I prefer archiving), messaging other people, and spending money.
Many AI tools that can act for you have a setting for this, often called "permissions" or "approvals." In many, you choose for each kind of action whether it can just do it or has to ask you first.
Use the setting, not the instructions you type in the chat. Typed instructions can get lost, like Summer thinks hers did. Her best guess is that the agent had too many things in its memory and lost track of her "confirm before acting" rule. A setting doesn't depend on the AI remembering. If your tool has no setting like this, don't connect it to anything you can't undo.
Important tip! An AI that reads your email or web pages can be tricked by what it reads. Someone can hide instructions in an email or on a website, and the AI may follow those instructions as if they came from you. It could send your private information to a stranger, delete your files, or buy something. So for any bot that reads email or the web, only connect the apps it needs. If it's sorting your email, it doesn't need your bank or your calendar. And double-check that everything you can't undo is set to ask first.
And know how to shut it down it before you start it, so you’re not running to your computer like Summer.
2. Plan for when it gets stuck
A bot that gets stuck does one of two things. It fakes its way through, or it stops partway.
My first bot that ran on its own was Morning Tea. Every morning, it sent me a news update. It was great at first, until one day I clicked a link it gave me and I got sent to an error page. The article didn't exist.
I'd asked Morning Tea for three to five interesting news stories a day about AI in the creator economy. And when it ran out of real news, it made some up to hit that number, with fake sources and all.
So tell every bot what to do when there's nothing to do. Ask for "up to five," instead of "three to five." And add a line to its instructions like "If there's nothing real to report, say so. Don't make things up."
My iPhone Duo bot was my other blunder. Pre-orders opened in the middle of the night, so I asked my AI to order one for me while I slept. (Spending money is on my ask-first list, so I approved that one order ahead of time.)
Before I went to bed, I asked if it had everything it needed. It said, "I got you. Don't worry about it. Just go to sleep."
I went to sleep, all warm and fuzzy, because it "got me." But when I woke up, I saw it had hit a sign-in page it couldn't get past and stalled out. Turns out it did not "got me." And I never got my iPhone Duo.
The first time a bot uses a new site or app, it often needs you there to help it sign in or set things up, even if it says it can do it. So always do a test run while you're awake. And don't take its word that it's ready. Make it show you.
3. Check its work, and don't let it grade its own homework
Even when a bot doesn't get stuck, the work can get sloppy.
Start by asking for a source link on everything it tells you. But remember, Morning Tea gave me links too. A link only helps if someone clicks it.
That someone doesn't have to be you. I have reviewer bots check my bots' work, and they catch most made-up stuff before it reaches me. A reviewer is a second AI whose only job is to check the first one's work.
How careful I get depends on how much is riding on it. For everyday jobs, my bot hands its work to a separate helper bot to check. (Some tools call this a "subagent.") It runs inside the same chat, but it starts fresh and only sees the work it's handed, so it isn't grading its own homework. If it's important, I paste the work into a brand-new chat myself. If it's VERY important, I have AIs from a few different companies triple check it.
Here's an example of what you could ask your reviewer. Attach or paste your bot's work right under it, and use an AI that can browse and search the web.
You're a fact-checker. Your only job is to check the work below. Treat nothing as true until you've seen proof.
1. List every fact in it: names, numbers, dates, quotes, events, and anything else stated as true.
2. Find proof for each one. Open the link it came with. If there's no link, or the link doesn't back it up, search for a reliable source yourself.
3. Match the details, not just the topic: same number, same date, same person, and the exact words for quotes.
4. Watch for weak proof: dead links, old news passed off as new, and sources that just repeat someone else's claim.
Sort every fact into one of these:
❌ WRONG: what's actually true, with a link.
⚠️ PARTLY RIGHT: what's off and the correct version, with a link.
❓ CAN'T CONFIRM: what you tried and why it didn't work.
✅ CONFIRMED: quote the line that backs it up, with the link.
List them in that order, problems first. End with one line: use it as is, fix it first, or start over.
Don't rewrite the work, and don't guess. If you're not sure, it goes under CAN'T CONFIRM.Even with a reviewer, read the first few runs yourself. After that, keep spot-checking every so often, and check a few of the reviewer's quotes against the page.
And expect some slop. AI gets sloppier the longer a job runs and the more it has to read at once. Facts that are hard to find online trip it up too. Creative work is the hardest to check. A reviewer can check a fact against its link, but it can't reliably tell you if a headline or image is good. So the more creative the job, the more fixing it'll need from you.
THE 30-SECOND VERSION
Let your AI do anything you can undo. In the tool's settings, make it ask you first before deleting, messaging people, or spending money. Don't rely on typed instructions for this.
If it reads email or the web, someone can trick it with hidden instructions. Give it only the apps it needs, and keep everything you can't undo on “ask first”.
Know how to stop it mid-job before you start it. Typing "stop" may not be enough.
Plan for when it gets stuck. Ask for "up to" a number, tell it what to do when there's nothing to report, and watch a test run while you're awake.
Ask for a source link on everything. Have a second AI check its work, with more checkers the more it matters. Read the first runs yourself.
Set the rules before you set it free. Then you can actually go to sleep.
me+machine.