Think of an AI agent like hiring a new employee versus booking a meeting. A meeting is where you ask questions and walk out with notes you still have to act on. An employee is where you hand over the outcome and they go run the whole thing. That’s the exact analogy the creator of this video uses to break down AI agents, and honestly, it made the whole concept click for me in a way nothing else has. He says his team runs hundreds of agents doing 92% of the work across his companies. I was skeptical until I saw how simple his framework actually is.
Let me map his analogy to the real steps.
🧠 The meeting vs. the employee
The original poster draws a hard line between chat and agents, and it comes down to who’s pulling on who.
- Chat pulls on you. You prompt, you wait, you copy and paste the answer somewhere useful. You’re still doing the work.
- An agent pushes on you. It runs the workflow, changes things, then checks in to confirm it landed right.
- The difference is the loop. Without a loop, it does the job once and stops. That’s automation. With a loop, it reviews itself and improves.
He breaks the agent’s “body parts” into an acronym he calls DATA: Diagnose (figures out the problem like a consultant), Assemble (builds the plan like an architect), Take action (executes), Assess (checks its own work and fixes what’s off).
With chat, you buy back your time. With an agent, you let go of whole areas.
🎯 Should you even hire this employee?
Before you build anything, this industry pro runs the task through his “rule of R”:
- Repetitive. Do you do this every week? If it’s a once-in-a-while thing, skip it.
- Rules-based. Same input, same output, every time? If the process is fuzzy, an agent will flail.
- Return on time. If the task takes 2 minutes but the agent takes 2 weeks to build, just keep doing the 2-minute task.
Fails the test? Stick with chat. This filter alone saves you from a pile of wasted weekends.
🏗️ The AGENT framework: your onboarding checklist
Here’s where the analogy really pays off. Building an agent is basically onboarding a new hire.
A — Aim for a specific outcome. When you hire someone, you don’t hand them a step-by-step script. You tell them what you need: grow the business, get more customers. Same here. Give it the why before the how, because it might find a better path than you would. His inbox example: “I need to spend less time managing my email inbox.” Notice there’s zero “how” in there.
Then write a definition of done. Not “handle my emails.” Instead: “Every morning at 9am the inbox is empty, replies are drafted in my voice, anything that needs me is flagged to the top, and nothing important slips.” If you can’t picture it done, the agent can’t hit it.
His advanced move is reverse prompting: state the result you want, then tell the AI to ask you the questions it needs to fill in the gaps. Let it build the plan.
G — Give it an identity. Out of the box, AI knows a little about everything and nothing specifically well. He cites a report where an airline’s support agents dropped from a 33% success rate to 11% after the rule books were stripped out. Same model, same task, three times worse, just because it forgot who it was.
His fix is three plain-English files:
- Soul file — how it behaves. Personality, quirks, tone, values.
- Identity file — who it is. Name, role, job, and the lanes it must never leave.
- User file — who it works for. Your goals, your priorities, your context.
Pro tip from him: don’t write these yourself. Feed the AI your aim, ask it to create all three files, and tell it to interview you first. It hands back something 99% done.
E — Equip it. Picture a genius sitting at a desk. The desk is the context window. On it: playbooks (your processes), identity files (its constitution), tools (logins and connectors), loops (its schedule). Under the desk: filing cabinets, aka memory. Pile too much on the desk and you get context rot, where it answers but isn’t clear or certain because it can’t find anything.
To capture your processes, he offers two routes. The camcorder method (record yourself doing the task while talking through it, then have AI turn it into a playbook) works but isn’t his favorite. The better move is reverse engineering from the source: connect the AI to your Gmail, have it read your last 50 sent emails, study your tone and sign-offs, and write your style guide. You’ve already done the work. Let it learn from your history.
N — Narrow the scope. One specialist per job. He has one agent that writes code and a separate one that reviews it. You wouldn’t ask your assistant to also run marketing and take sales calls, so don’t build a mega-agent. Instead, build a manager agent whose only job is coordinating sub-agents. His is named Kai. He talks to Kai, Kai talks to everyone else.
He also drops a model-picking tip: Haiku for simple high-volume sorting, Sonnet for day-to-day work, Opus for complex reasoning and management, Fable for long-running orchestration. He once ran a full code refactor on Haiku for $1.50 that would’ve cost about $150 on Opus. His inbox agent runs on Sonnet because a defined process doesn’t need genius-level horsepower.
T — Trust in stages. This is the scary part, and he’s blunt about it: you don’t hand over the car keys on day one. His ladder:
- Set guardrails in the identity files. Can it spend money? Send emails? Drafts only?
- Approve everything at first. “Show me what you’d do.”
- Loosen the leash as it earns it.
- Give it a heartbeat. A real schedule, running on its own.
When he showed the inbox agent to his executive assistant, she thought she was out of a job. It freed her up for higher-level work instead.
⚡ Where to start today
Run your week through the rule of R and find one task that’s repetitive, rules-based, and worth the build time. Inbox triage is the obvious first hire. His closing pro tip is almost too good: hand this video’s link to your AI and tell it to build the agent using everything in it.
The full video walks through every prompt and example in detail. Worth watching end to end.