OK so the title is a lie. Sort of. Bear with me.
I make about $5,000 a month right now from a handful of small AI agents I built and maintain for small-business clients. The agents do the work. I check in once a week to make sure nothing's broken. So "sit back" is true in the sense that I don't have to be at my keyboard for the income to come in.
It is not true in the sense that the eight months before this looked anything like sitting back. The eight months before this looked like figuring out what an AI agent even was, building things nobody had asked for, charging $100 for something I should have charged $1,500 for, and DMing strangers on LinkedIn with messages that read like ChatGPT wrote them.
Anyway. That's the disclaimer. The income is real. The "sit back" is doing some heavy lifting. Here's how I actually got here.
If I could go back and give past-me one piece of advice for those eight months, it would be this. Stop watching YouTube tutorials. Pay for a structured course. Finish it. I lost months because every tutorial contradicted the next. Different stack. Different pricing. Different definition of what an AI agent even was. One paid course with daily fifteen-minute lessons and I would have shipped my first paying client agent inside a week. YouTube is fine once you already know what you are looking for. It is hostile when you do not.
How I started, and how dumb it was
Late last year I was scrolling Twitter and kept seeing people post things like "agents are going to change everything" and "I just built an agent that books my meetings and writes my emails." I had no idea what they were talking about. I knew what ChatGPT was. I'd used Claude once. I didn't know there was a meaningful difference between "asking Claude a question" and "building an agent."
So I went down the rabbit hole. Spent maybe two weeks reading threads, watching YouTube, and most usefully, just opening Claude and trying to make it do small tasks. Sort my email by importance. Read a PDF and summarize the financial table inside. Draft three versions of a follow-up message to a colleague.
That stuff isn't agentic yet. It's just structured AI use. But it taught me how to talk to the model, what it was good at, what it would hallucinate. By the time I started actually wiring things up with tools (web search, document reading, calendar access), I already had the language to make Claude do what I wanted.
I built five or six tiny agents for myself. One that pulled my Stripe invoices into a Notion doc every morning. One that drafted LinkedIn replies in my voice. One that scanned my email for things I'd forgotten to reply to.
I was showing one of these to a friend over beers. He said "Dude. This is wild. You could sell these."
I said "to who? And why would anyone pay for it?"
That was the whole start of the business.
The cold DM disaster
He told me to DM small-business owners on LinkedIn. So I did. My first round of messages was about four paragraphs each. I explained what an AI agent was, how mine worked, what value it would provide, and ended with "would love to chat."
Zero replies out of about fifty.
I asked him what was wrong. He said "Nobody reads four paragraphs on LinkedIn. Make it three sentences." So I rewrote them. Still zero replies, but less time wasted per send.
I was about to quit cold outreach entirely when something else happened. I'd posted a video on Reddit (in a small-business sub) showing one of the agents I'd built. Just a 90-second screen recording of me asking it a question and watching it complete a task. The caption was basically "I built this for myself, thought someone might find it cool."
A guy who ran a small e-commerce store DMed me. "Can you make something similar for me? My customer service emails are eating my week."
That was the first client.
The $100 mistake
I had no idea what to charge. He was nice, the request was small (a Claude-powered email triage agent that drafted responses he'd review before sending), and I was so happy to have my first real ask that I told him "$100, one-time, to build it."
He paid immediately. Probably because $100 was insane for what I built.
I learned later, by accident, from another conversation, that similar agents in his industry were going for $500 to $1,500 a month as ongoing services. I had built it as a one-time deliverable for a price that was approximately 1/15th of what it was worth.
I didn't go back and renegotiate because that's not how I'm wired. But I changed the offer for client #2. Same agent, customized for their business, $500 setup + $300/month for hosting, updates, and tweaks.
They said yes.
What "sit back" actually means
Today I have 12 agents running for 9 clients. Some clients have multiple agents (one for support, one for proposals, one for lead qualification). The cheapest is $250/mo, the most expensive is $1,200/mo. Average is around $560. Total monthly is roughly $5,000.
The "sit back" part is technically accurate. The agents run on their own. Claude does the work. I check the logs every Monday morning to make sure nothing's drifted, fix anything broken, and respond to client messages.
The total time I spend on the maintenance side is maybe 4 to 6 hours a week. The rest is finding more clients, which is the only real lever I have left.
What I do not sit back on, ever.
- The first month of a new client. I'm in there constantly, watching what the agent gets wrong, retraining the prompts, building in fallbacks. The "good" version of an agent takes about 30 hours of work over 3 to 4 weeks before it's stable enough to run unattended.
- Anything client-facing. People who pay $500 to $1,200 a month want to feel like a person is on the other end. There is. It just isn't doing the agent's job, it's doing the relationship's job.
- Updates when Claude releases something new. Last time the API context window shifted, I spent a weekend retesting every agent to make sure nothing regressed.
So "sit back" is closer to "I built a small portfolio of products that work without my constant attention." Like owning a rental property, except cheaper, faster, and the property is software.
The pipeline now
I pivoted from cold DM almost entirely to inbound. Most of my leads come from posting builds on LinkedIn now (I made the brand jump from Reddit to LinkedIn around month 5), and a smaller number come from referrals from existing clients. I haven't sent a cold DM in two months. Posting cases of "here's a real agent I built, here's what it does, here's the rough effect on the business" works way better than DMing strangers about agents in the abstract.
I learned the prompt structure for all of this from a mix of stuff. Twitter threads. A short course I did on Mindwand early on, which was 15-minute daily lessons and taught me how to structure conversations with Claude in a way I could repeat. Coursera and the bigger $497 creator courses are other options that probably also work. The point of any of these is to take the chaos of "I'm just typing words at Claude" and turn it into a system you can teach a tool to follow. I picked Mindwand because it was the cheapest one and I finished it. Format matters more than brand.
The honest tip
If you're going to do this, the bottleneck is not Claude. Claude is the easy half of the problem. The bottleneck is identifying a specific, painful, repeated task at a specific business and building exactly the thing that solves it.
I spent the first two months building agents nobody had asked for, then trying to sell them. None of those sold. The agents that sold were the ones where I asked a business "what do you do every week that you hate" and built a thing that did that.
Inbox triage. Lead qualification. Proposal first drafts. Content repurposing. Customer FAQ answering. Small, repeated, painful, expensive in time. Those are the agents that print money. The agents you build because Claude can technically do them, but that nobody asked for, don't.
That's the whole game. Sit back is the destination. Don't confuse it with the starting position.