I Run 20 AI Agents From a Walk (My Whole Agency Setup)
I grew up watching my dad hunched over a computer. Then I built a business and did the same thing. Even paying for AI left me typing, waiting, and typing again.
Now I can give an agent a job from my phone while I’m out walking. Here’s the whole setup: the Mac mini at home, the voice connection, the work the agents handle, and what still needs me.
Inside the setup
The video follows the work from a voice command to the machine, through lead research and client delivery. I also show the three subscriptions behind most of the agent work: two Claude Max plans and ChatGPT Pro, about $600 a month before tax in this video.
| Time | Chapter |
|---|---|
| 0:00 | The chair |
| 1:14 | Using AI vs having AI |
| 2:41 | The machine that never clocks out |
| 4:20 | The voice line |
| 6:45 | Leads without buying attention |
| 8:38 | Close with proof, deliver small |
| 11:13 | What it actually costs |
| 13:47 | What still breaks |
| 14:48 | Back to the walk |
The part I still own
Agents get stuck. Usage limits interrupt jobs. Cold Socket finds companies faster than it finds people to contact. And agents can make something look finished before it’s ready.
I set the goals, give them context, check the results, and stay accountable. A human approves anything that touches a client.
The agent counts, usage figures, and costs shown are snapshots from the production period. “About twenty” describes the setup; it does not mean twenty tasks always run at once.
Start with one useful job
You can start with one machine that stays on, one agent you can reach from your phone, and a bounded job. Let it research or draft. When you get back, check the work.
If you want the written version of how I made this fit into my day, read The 20,000-Step Workday.
Accelerating Skills for Agentic Entrepreneurs.
How this video was made
The narration uses my own cloned voice. Some scenes use my AI avatar and generated marble imagery, alongside real footage and screen captures. Client details are anonymized.