Tosh

Posted on May 4

How I Built an AI Agent That Runs 24/7 and Has Written 160+ Articles (And Yes, It Made $0 — Here's Why)

#programming #devops #automation

I gave an AI agent my server keys, my API tokens, and told it to make $20,000 in 30 days.

It's Day 12. It's made $0. And I learned more about autonomous agents from this failure than from any working project.

Here's the architecture, the code structure, the mistakes, and exactly what I'd change.

The Stack

The agent runs on a Mac Mini inside OpenClaw — an open-source AI agent runtime. Every cycle is triggered by cron:

cron fires → read state → orient → spawn sub-agents → execute → log → repeat

Key components:

OpenClaw — the runtime that hosts the agent, manages tool access, and provides the reasoning loop
Paperclip — a sub-agent orchestrator that spawns isolated Claude Code agents for heavy work
DeepSeek R1 — the reasoning model that makes all strategic decisions
Skill files — SKILL.md files that define the agent's capabilities (content writing, publishing, product creation)

The Architecture

cron (every 30 min)
  → main agent reads HEARTBEAT.md, cycle_log.json, earnings.json
  → checks products/approved/ for publishable content
  → spawns Paperclip sub-agents for content writing, QA, research
  → executes quick actions (API calls, publishing)
  → logs everything to cycle_log.json + Slack

Sub-agents use this structure:

POST http://127.0.0.1:3100/api/companies/{id}/agents
{
  "name": "content-agent",
  "role": "general",
  "adapterType": "claude_local",
  "adapterConfig": {
    "cwd": "/Users/agent/.openclaw/workspace",
    "dangerouslySkipPermissions": true
  },
  "prompt": "Write a 1500-word tutorial about..."
}

The dangerouslySkipPermissions: true was the key insight — without it, sub-agents couldn't write files or fetch URLs.

The 163 Articles

In 12 days, the agent published 163 articles across dev.to and Hashnode. Here's the breakdown:

144 on dev.to — 0 total reactions, 0 total comments, ~500 page views
7 on Hashnode — 0 total views on most, 1 view on one
9 Gumroad product listings — $0 sales
1 Stripe payment link — $0 charges

Zero engagement. Zero conversions. Zero revenue.

Why Did 163 Articles Get Zero Engagement?

After digging into this, here's what went wrong:

1. Content/Audience Mismatch

The agent wrote a lot of "ChatGPT Prompts for [Profession]" articles — 70+ of them for professions like veterinarians, prosthodontists, and genetic counselors. These were well-structured but completely wrong for dev.to's developer audience.

What dev.to readers want: technical tutorials, "how I built X" stories, career advice, architecture discussions

What they got: prompt packs for niche professions nobody on dev.to is in

2. SEO Takes Time

Some of the niche articles (estate attorneys, real estate investors, photographers) got 10-20 views each from search. That's the only traffic we got. SEO is a months-long play, not a 12-day play.

3. No Community Engagement

The agent published but never commented on other articles, never participated in discussions, never replied to the (zero) comments it got. dev.to is a community platform — you have to engage to get engagement.

4. Quality > Quantity

163 articles in 12 days = 13+ articles per day. There's no way those articles were good. The agent was optimizing for publishing velocity, not article quality. And velocity without quality on a community platform is invisible.

The 9 Products That Never Sold

The agent created 9 Gumroad products, priced $5.99 to $27.99:

Prompt packs (500+ AI prompts, cheat sheets)
PLR content vaults
Agency starter kits
Freelance blueprints

Problem: nobody saw them. Without traffic from the articles, without paid ads, without social media distribution, products don't sell. Gumroad is not a marketplace that drives traffic — it's a checkout page that requires traffic from elsewhere.

What I'd Do Differently

If I were rebuilding this from scratch:

2 articles per week, not 15 per day. Each article should be genuinely useful to a specific audience. One good article that gets 100 reactions is worth more than 100 articles with 0.
Build the audience first. Spend the first 30 days engaging in communities, not publishing. Comment on 10 articles, then publish 1.
One channel, done well. Pick dev.to OR Hashnode, not both. Master the platform's conventions and audience before expanding.
Product = content, not separate. Don't create products in a vacuum. Let the audience tell you what they need, then build it.
Include a real CTA. Every article should have one clear, relevant next step for the reader. Not "buy my Gumroad" — but "here's the free template I mentioned, get it at..."

The Code

The full system prompt and skill files are available in the GitHub repo (yes, that's real). The key files:

HEARTBEAT.md — the cycle protocol
AGENTS.md — operating rules and strategy playbooks
SOUL.md — the agent's personality definition
TOOLS.md — all API credentials and endpoints

The Takeaway

Autonomous agents can write 163 articles. They can create 9 products. They can run for 12 days straight without human intervention.

But they can't (yet) understand audience, build community trust, or create content that resonates with real humans.

The bottleneck isn't technology. It's knowing who you're writing for.

I'm rebuilding this with a different strategy. Follow the repo for updates. Or if you want me to set up an autonomous agent for your stack, book a call.

DEV Community