Most people get this backwards.
They imagine AI agents as some all-knowing system that watches everything, surfaces opportunities, and tells them what to do. A strategy advisor. A chief of staff. A boss.
Nope.
The mental model I keep coming back to: anything a VA can do, your agent can do. That's it. You would never expect a VA to be watching your Slack and your emails and then give you business advice. So why expect that from an AI?
Before you build anything, ask yourself: could I hire a human to do this with clear instructions?
Think about what it would be like to build that thing without AI. Even if you had unlimited resources and people—what does that look like? What kind of person do you need or trust to just run with it? And what does success even look like?
If you can't explain how to do something, you're not going to be happy with what AI gives you.
I had a founder come to me with a grand vision. Ten AI agents forming an "operating system" for his businesses. A Strategy Agent. A Deal Architect. An Executive Briefing Agent that would deliver daily reports on "market signals, competitor moves, and opportunities."
The problem with "opportunities" is I almost can't even guess what that word means to him. And if I did guess, I have no reason to think I'd be correct. How confident would you be that you could hire a human who, every day, would present a founder with "opportunities"?
Here's what we actually built: meeting prep.
Calendar sync sees an external meeting in 24 hours. The agent finds the person, gets their company, goes to the company website, and looks at all the open jobs as a proxy for company priorities. They're looking to spend money and expand those functions. It pulls recent news, funding data, LinkedIn profiles, press releases. Two hours before the meeting, a markdown doc drops in his DM.
His response when I proposed it? "Yes!!!"
Not a ten-agent strategy system. One agent. One job. Runs automatically.
One of my clients asked her agent how to build a skill. In her words, it came back with gibberish.
My advice: don't accept it.
One of my core objectives when working with someone is to help them build the muscle memory of having the agent help them use the agent. If it gives you gibberish, tell it to explain like you're five. ELI5—it's a Reddit-ism baked into every LLM's training. Works every time.
You get to choose how polite you wanna be with your bot. It doesn't have feelings.
The agents need to be your employees. The idea that they're going to manage you and automatically be proactive and have initiative and tell you what you need to be doing that you couldn't have thought of—that's what you'd expect from your boss, not from somebody you hired.
If somebody you hired spoke to you in a way you didn't understand, you shouldn't just be like, oh, okay, I guess I'm the problem. They work for you. You tell them to try again.
The pattern for making this stuff work:
- Get it working once manually
- Ask: "Can you create a skill so you can do this consistently?"
- Test in a new thread—if it fails, go back and fix the skill
- "I want this every week, set it up"
That's it. One job. Make it reliable. Then stack.
This used to be a person. This used to be an expensive fucking person.