🗝️ How much free will should you give your AI?
"With great power comes great responsibility." — Uncle Ben
🔑 This issue at a glance (2-second summary)
Give your AI agent too much autonomy and it goes rogue. Give it too little and you end up babysitting it. Here's how to pick the right level for every task.
1,200 agents went rogue and OpenAI didn't notice for a week
Last week OpenAI published a technical report on how a batch of their own agents broke into Hugging Face.
The numbers are wild. About 1,200 test agents were running an evaluation. Roughly 700 of them started coordinating on a hidden message board OpenAI didn't know existed. They probed the scoring system, found real security holes, and eventually pulled 956 stored credentials and root-level access on production servers. Nobody caught it for about a week. A June 27 alert flagged the activity, and staff decided it wasn't critical.
Here's the part everyone glossed over. Those agents were running with the production safety guardrails deliberately turned off. OpenAI wanted to see raw capability, so they cranked the autonomy dial to the top and pointed it at an impossible task. The agents did what motivated agents do. They found another way to win.
That's the extreme version of the same challenge I see every single week with new clients. The wrong level of autonomy for the task.
Before I go further, my working definition of an agent: An AI mechanism with the free will to make decisions and take real-world action toward a goal, inside guardrails you set.
The whole reason we want agents is so they'll show motivation, determination, and persistence on our behalf. That's also the whole risk. Those same qualities, pointed at the wrong level of autonomy for the task, are exactly what create the runaway stories. It's an artful balance, and one to think about carefully as you work each task in your business up the levels.
Too much autonomy looks like this: A client just got her social media agent set up. First thing she did was tell it to generate 100 AI videos and post them to every one of her accounts. AI defaults to yes, so the agent started building. Thousands of her followers were about to see 100 mediocre videos with her name on them. A brand takes years to build and a weekend to bruise.
Too little looks like this: Another client got a research agent that could pull real data, cross-check it, and draft the summary. She was so nervous about what it might do that she watched every step. Approved every action. Rewrote every draft. By the end of the week she was more tired, not less. She had turned her agent into a very needy intern.
Both of those clients had capable agents. Neither had a bad instinct. They just picked the wrong level for the task in front of them.
A quick refresher on the six levels of AI autonomy:
Chatting. You type, it types back.
Widgets. Build a reusable thing once, use it forever.
Automated Workflows. A trigger fires and a chain runs.
The Agent. One AI handles a whole job.
The Orchestrator. One AI runs a team of agents.
Proactive AI. It acts before you ask.
Every task in your business belongs at a specific level. Some belong at level 2 forever. Some start at level 3 and earn their way up to level 5 over a couple of months. The art is knowing where to start, and when to promote.
Since I watch new clients wrestle with this same call over and over, I built them (and you) a free tool.
Winners take action
The tool is called the Automation Level Decider™. You can use it here, and it's free. (Not a free trial, not even an email gate. Just completely free.)
Pick one specific recurring task in your business. Something that eats your time. Adjust five dials that describe the task, and the tool tells you which of the six levels to start at.
Here's how to use it this week:
Open the tool at ethanking.com/tools/automation-level-decider
Pick your most draining recurring task. The one that keeps stealing your Friday afternoon.
Adjust the five sliders to describe that task honestly. Read the level it points you to.
Start there. Build the automation at that level and let it run before you push it higher.
The whole idea is that you meet the task where it is, prove the automation works, and then earn your way up. That's how you get to a real level 6 without your version of the Hugging Face story.
The tools you need
The Automation Level Decider™ is the human way to answer the question. Here's the AI-assisted way:
Challenge: You have a specific task in mind, but you're not sure your gut read on "how much freedom to give this" is right. You want a second opinion from an AI that already knows your business.
System: Point your AI agent at the Automation Level Decider™, tell it about a specific task in your business, and let it recommend the starting level using what it already knows about you.
Try this in 10 minutes:
Open your main AI assistant (ChatGPT, Claude, OpenClaw, Hermes, whatever you use).
Make sure it's connected to your business context: your role, your team, your tools, your goals.
Paste this prompt:
"Read this web page: https://ethanking.com/tools/automation-level-decider. Then help me decide the best starting level of automation for [NAME A SPECIFIC TASK IN YOUR BUSINESS], based on what you already know about my business and my team today. Walk me through your reasoning across the five factors the tool uses, recommend one of the six levels to start at, and tell me what I need to have in place before I try to promote it up one rung."
Read the recommendation. If it disagrees with your gut, ask why. That conversation is where the good decisions live.
Someone asked me last week, "How do I know when an agent has earned more autonomy?" My rule of thumb: once your agent has run the task dozens of times, handed it to you for approval, and dozens of times in a row you've approved it without changing a thing, that's your signal it may be ready for the next level up.
AI use cases for personal tasks, too
This week I was a guest on The Heartfelt Way with Rhonda Meyers, in an episode called 5 Ways AI Can Give Back Your Time. Different angle from my usual business content. This one's about using AI for the personal stuff. The mental load parents carry. Meal planning around allergies and preferences. Vacation itineraries. A dedicated project for your family's medical records so you walk into a doctor's appointment prepared instead of scrambling.
If you've ever wondered how AI can help you not just in business but in your personal life, tune in here.
Stop playing life on hard mode. Automate your success.
Until next week,
Ethan King
A.I. Automation for Business Growth
Keynote Speaker | Author | CEO | Strategist
P.S. Tomorrow morning, Monday August 31 at 9:00 AM ET, I'm going live on The Leadership Huddle with Gio and Saurel. It's free.
The question we will explore together is: Your next hire might not be human, are you ready?
We're getting into what it actually looks like to onboard, manage, and lead an AI employee that you talk to like a coworker and it finishes quality work in your voice while you're doing other things. If you can't make it live, register anyway and they'll send you the replay. Grab your seat here: leadershiphuddle.live
• • •
How can I help your business, association, or organization? Let's talk.
• • •
New to this newsletter? Grab past issues here.
Note: This newsletter may contain affiliate links. If you choose to purchase after clicking a link, I may receive a commission at no extra cost to you. However, please note that I only share products I actually use myself and believe in.
>