How to Start Building AI Agents If You Have a Day Job

The cleanest moment in any agent build is the one before you start, when the path looks impossible.

I had that moment 18 months ago. Twenty years selling cars in DFW. No tech background, no computer science degree. Three businesses I was trying to start in evenings and lunch breaks, and a phone full of half-finished notes from calls I had let slip. The work was real. The capacity to do it was not.

What broke the loop was not a course. It was building one small useful thing, then another, then another, on a stack that turned out to cost nothing to start. Eighteen months later that path has 22 production agents on it, running across 4 businesses, on less than $50 a month in hosting. Pre-revenue, still selling cars on the weekends.

This is the sequence I would give to anyone starting today with a day job and a real workload.

Can I really build AI agents without a tech background?

Yes, with one honest caveat.

The threshold for building useful production agents has dropped much faster than the average career has updated to know that. You do not need to write code from scratch. You need to be able to follow a build, copy and modify what a model gives you, and read enough of the result to fix it when it breaks. That set of skills is closer to using a complicated spreadsheet than to traditional software engineering.

The caveat is that the no-code agent platforms most operators try first are usually the wrong starting point. They promise zero code and they deliver a wall the moment your real workflow touches a tool they did not anticipate. The agents that hold up over time are the ones built on small custom stacks where you can fix the wall when you hit it.

That sounds harder than it is. Modern model assistants will write most of the code for you. Your job is to know what you are trying to build and to read what comes back well enough to spot when it is wrong. Both of those skills come from doing one small build, not from reading about a hundred.

How much time per week do I need to start?

The real answer is unromantic. Five to eight hours of focused, uninterrupted evening or lunch-break work in a typical week. That is enough to ship a working v1 of a single useful agent in two to three weeks.

The math only works if those hours are protected. Thirty minutes split across five interrupted days produces nothing. One sit-down of two to three hours, twice or three times a week, produces a real build.

The discipline that made this work for me was a simple rule about which builds count. A build counts when:

• It runs on its own, without me triggering it.
• It produces output I can look at the next morning.
• It survives a real day of usage before I declare it done.

A build that requires me to keep tweaking it on the dealership floor is not done. A build that survives a Tuesday is done. That threshold is what kept me from going down two-week tangents that produced nothing usable at the end.

What is the first agent I should build?

Build the agent that handles the workflow you hate doing on your phone in line at the drive-thru. That is the highest-leverage starting point.

The category does not matter as much as the criteria. The starting agent should be:

• Small. It does one thing.
• Painful. You skip it now and it costs you something every time you skip it.
• Daily. You do this work, or fail to, five days a week or more.

For me it was a voice flag system that captured follow-ups the moment I finished a call. For a service writer it might be a triage script that classifies incoming voicemails. For an SMB owner it might be a script that pulls last night's revenue from your platform and emails you a one-line summary at 6 AM. For an operations manager it might be a one-tap script that closes a routine ticket and logs the outcome.

The pattern across all of these is the same. A small piece of automation that lives between conversations and prevents a class of dropped balls.

The workflow you hate

→

One small build that handles it

→

Run it for a week

→

Fix what breaks

→

Ship the next one

What should the first build look like, end to end?

Five steps that have held across every agent I shipped.

Step one. Pick the workflow. Write it down in one sentence. "When a customer calls and asks about service, I forget to follow up by Thursday." Be specific. Specific workflows produce specific agents. Vague workflows produce vague agents.

Step two. Pick the smallest possible agent that handles it. Not the perfect agent. Not the agent that solves five problems. The smallest one. A voice mic that commits a flag is smaller than a full triage system. A bash script that closes a ticket is smaller than a CrewAI multi-agent setup. Pick the smaller one.

Step three. Build it on the stack that costs nothing to start. Vercel free for the web layer. Railway hobby tier for any backend service. GitHub free for code and audit trail. Pay-as-you-go on Anthropic or OpenAI for the model. If a tutorial tells you to start with a $400-a-month enterprise tier, find a different tutorial.

Step four. Use it for a week without changing it. This is the step most people skip. The build is not done at v1; it is done after a week of real usage in a real workload without you needing to fix it. That week tells you which features actually matter and which ones were imagined.

Step five. Write down what broke. Specifically. With numbers when you have them. The next agent gets built on the lessons of the first one. Without the writeup, the lessons go away.

What should you not do?

Three traps.

Do not start with a chatbot. A chatbot is a conversation, and most operator workflows do not need a conversation. They need a capture, a route, or a close. Build those first.

Do not start with a no-code platform that promises everything. Those platforms produce a working demo and a dead end. The wall is real. You can either hit it now and learn what you need, or hit it in six months after you have spent real time on a build you cannot extend.

Do not buy the course first. The course is the long way around. The short way is to pick one painful workflow, follow a five-step build with a model assistant, and ship one agent in two weeks. If you do that and the work is enjoyable, you will know whether to go further. The course will be more useful once you have a build to compare it against.

The build that takes 14 hours of evenings and lunches and works on Tuesday is more valuable than the 80-hour course that produces nothing.

What is next once you have one working agent?

Build the second one.

Pick a different workflow you hate. Apply the same five steps. Ship it. The compound effect is the part nobody talks about, and it is real. Two agents in a quarter. Six in a year. Twenty-two in 18 months if you are lucky and the workflows keep showing up.

The reason this works at the operator level and not at the consultant or career-change level is the same reason it stays unromantic. You are not trying to be a developer. You are trying to remove the next dropped ball from your week. Each agent is one less ball.

If you want the full breakdown of how I went from zero agents to 22 in 18 months, what each one cost to build, and what each one produced last week, the playbook is in the 10 agents lead magnet. The community where we ship in public is at skool.com/agent-empire-4291. Free.

While I sell cars for a living.

Michael