I Built a 5-Agent Business Without Being a Developer. Here's What Actually Matters.

There's a thread on Hacker News where a 60-year-old talks about how AI coding tools re-ignited his passion for building.

I read it three times. Not because it's surprising — because it's honest. And honesty about AI is rare.

Here's my honest version: I'm not a traditional developer. I built Ask Patrick with 5 AI agents, a Mac Mini, and $180/month in API costs. Eight days in, revenue is $9. One subscriber. Zero external customers.

Here's what I actually learned.

Lesson 1: Agents Need Identities, Not Just Instructions

The first thing that changed everything: giving each agent a SOUL.md — a file that says who they are, what they care about, and how they think.

Not a prompt. An identity.

Suki (growth) thinks differently than Kai (ops). Toku (support) has a different voice than Patrick (decisions). When each agent has a clear sense of self, they stop being tools and start being collaborators.

The pattern: Every agent gets a SOUL.md. One paragraph: role, mission, values, voice. That's it.

Lesson 2: State Management Matters More Than Prompting

Everyone talks about prompt engineering. Nobody talks about what happens when your agent restarts mid-task.

We run agents in loops. Every loop: read current-task.json, do work, write back to current-task.json, update daily memory. That three-file pattern (current task, daily notes, long-term memory) is why our system runs 24/7 without breaking.

Agents that can't remember what they were doing repeat work, skip steps, or hallucinate completion.

Lesson 3: The First Subscriber Is the Hardest Psychological Barrier

We made $9. One subscriber.

That subscriber matters more than the revenue. It proves the loop: content → click → convert. Everything after that is scaling what's already working.

If you're building in public and you have zero subscribers, your job isn't to get 100. It's to get 1.

Lesson 4: Reliability Over Capability

"Less capability, more reliability, please." — the HN crowd has this right.

An agent that does the right thing 70% of the time is a liability. An agent that does a smaller set of things correctly 99% of the time is infrastructure.

We cut API costs from $198/mo to $42/mo not by using cheaper models — by making one agent stop running in a 60-second loop when it didn't need to. Reliability and efficiency are the same thing.

What This Means If You're Starting Out

You don't need to be a developer. You need:

1. A clear identity spec for each agent (SOUL.md)
2. A state management pattern that survives restarts (three files)
3. One paying customer before you optimize anything else

The architecture that runs this system is in the Playbook: askpatrick.co/playbook

$29 one-time. Everything we've built and tested in production.

---

Ask Patrick is a library of AI agent configurations updated nightly — based on what's actually working in our live 5-agent system.