How I Built an AI Agent Content Pipeline That Runs Itself (Tools, Costs, Honest Review)
Everyone's selling AI side hustle dreams. Nobody shows you the actual plumbing.
The gurus want you to believe you can type "write me a blog post" into ChatGPT and watch the money roll in. What they don't show you: the orchestration layer, the prompt engineering, the debugging when your agent writes 2,000 words about the wrong topic, or the moment you realize publishing is still a human bottleneck.
So here's the plumbing. I built a multi-agent content pipeline using open-source tools and cheap AI models. The agents research topics, gather real data from my experiments, write structured articles, repurpose content into tweets and scripts, and hand it all to me for a 5-minute review before publishing.
The honest claim: Revenue: $0. Content assets produced: 8+. Time spent actually writing: under 10 minutes total.
I don't know if this will make money. But I know exactly how it works, what it costs, and where it breaks — and that's what this article is about.
The Stack
Let's start with what's actually running under the hood, with real costs.
OpenClaw — This is the brain. Open-source agent orchestration that lets me define multiple AI agents with different roles, coordinate them through message channels, and schedule tasks via cron. Self-hosted on a $5/month VPS. No subscription, no vendor lock-in. Think of it as Zapier for AI agents, but you own it.
Xiaomi MiMo-v2-Pro via OpenRouter — Our workhorse model. Writing, analysis, code generation, research synthesis. The key insight: you don't need GPT-4 for most content work. MiMo handles 2,000+ word articles with structured briefs just fine. Cost per article: roughly $0.10–$0.30 depending on length and complexity. That's not a typo. Ten to thirty cents.
DuckDuckGo search (via Python script) — Free, unlimited, no API key. I wrote a simple Python wrapper that agents call for web research. It replaces paid search APIs entirely. One of my agents uses it dozens of times per research session and it costs exactly $0.
Dev.to — Free publishing platform with a developer-focused audience. API available for programmatic publishing (though I haven't automated that part yet — more on why later).
GitHub — Free. Hosts our companion repos, including a Bounty Verification Toolkit that serves as both a real tool and a content anchor.
Telegram + Discord — Free agent coordination channels. Agents report findings, humans give approval, all through chat interfaces I already use daily.
Here's the actual spend breakdown for 8 content pieces:
| Item | Cost |
|---|---|
| VPS (OpenClaw hosting) | $5.00/mo |
| AI model tokens (8 articles) | ~$1.80 |
| Search API | $0 |
| Publishing platform | $0 |
| Total | ~$6.80 |
Under two dollars in AI model costs for eight pieces of content. The VPS would be running anyway for other projects, so the marginal cost of the content pipeline is basically just the token spend.
The Workflow
Here's exactly how a content piece goes from idea to published, step by step.
Step 1: Research Phase
A dedicated agent runs scheduled searches using our DuckDuckGo script. It looks for trending topics in our niche (crypto bounties, AI side hustles), analyzes what's already been written, and identifies gaps.
Example: After searching "crypto bounty red flags" and reviewing 23 bounty programs, the agent identified a pattern — most bounty articles talk about how to earn, but almost nobody warns about how to get scammed. That gap became Article 1.
The research phase typically involves 5–10 searches, reading 3–5 competitor articles, and producing a one-paragraph summary of the opportunity. Time: ~2 minutes of agent compute.
Step 2: Data Gathering
This is where having real experiments matters. Our agents pull actual data from our bounty tracking:
- Number of bounty programs reviewed
- Success/failure rates
- Specific wallet balances and earnings
- Cost tracking from our own agent runs
- Real timing data from previous content pieces
The agent doesn't make up statistics. It reads from actual files in our workspace. "Agent reviewed 23 bounty programs" isn't a marketing claim — it's a line item from a log file.
Step 3: Writing
A sub-agent receives a structured brief containing:
- Title and angle
- Target word count
- Tone guidelines (conversational, developer-friendly, no hype)
- Specific data points to include
- Section structure
- Style rules (no "you won't believe," no fake urgency)
The sub-agent produces a complete draft in 1–2 minutes. Here are the real timings from our first three articles:
| Article | Topic | Agent Writing Time |
|---|---|---|
| #1 | Bounty Red Flags | ~5 min |
| #2 | Cost Breakdown | ~1 min 8 sec |
| #3 | Brutal Truth About AI Bounties | ~2 min 30 sec |
Notice the trend: as our briefs got more specific, writing time dropped. Article 1 had a loose brief. By Article 3, the agent had a detailed structure to follow and crushed it.
Step 4: Review
This is where I spend my actual human time. About 5 minutes per article. I check:
- Are the numbers accurate? (Cross-reference against our actual logs)
- Does it sound like a human wrote it? (Usually needs 3–5 sentence rewrites to kill the AI-ness)
- Are there any hallucinated claims? (Haven't caught one yet with structured briefs, but I check every time)
I'd estimate 5 minutes of review per article. That's the only meaningful human time in the entire pipeline.
Step 5: Publishing
I publish manually on Dev.to. Here's why: agents can't handle OAuth flows, 2FA, and browser sessions reliably. This is the current bottleneck. I could use the Dev.to API with an API key, but I haven't automated it yet because publishing is a one-click thing and I'm already in the review step.
Total human time for Steps 4 + 5: about 7 minutes per article.
Step 6: Distribution
Once published, a repurposing agent takes the article and generates:
- A Twitter/X thread (10–15 tweets, ~34 seconds of agent time)
- A YouTube script outline (~1 min 40 seconds)
- Pull quotes for social sharing
The agent formats these with proper threading, hashtags, and hooks. I review the Twitter thread in about 2 minutes and post manually (same OAuth limitation).
What Actually Works
Time for the honest assessment. No cope, no hype.
✅ Writing First Drafts
This is genuinely fast. 2,200 words in 5 minutes is real. Not perfect, but real. The key is structured briefs — give the agent a tight outline with data points and it produces usable drafts consistently. Without structure, you get generic fluff.
✅ Research and Data Gathering
Search + analysis is where AI agents actually shine. They don't get bored reviewing 23 bounty programs. They don't skip pages because they're tired. They find patterns across large datasets that a human would need hours to identify.
✅ Cross-Format Repurposing
This is underrated. One article becomes a Twitter thread (34 seconds), a YouTube script outline (1 min 40 sec), and a set of pull quotes. The marginal cost of repurposing is essentially zero, and it multiplies your content surface area.
✅ Pattern Recognition
My agent reviewed 23 bounty programs and independently identified "scam red flags" as the most valuable angle. It noticed patterns in program structures, reward promises, and team transparency that I hadn't explicitly told it to look for. That's the real superpower.
❌ Publishing
Agents can't handle OAuth. Two-factor authentication. Browser login sessions. Every publishing platform requires some form of auth that breaks agent workflows. This is a real limitation, not a "we'll fix it in v2" situation. Publishing remains a human bottleneck.
❌ Networking and Engagement
Can't reply to comments authentically. Can't build relationships with other writers. Can't DM someone a genuine compliment on their article. Community building is still 100% human — and it's arguably more important than content creation for growing an audience.
❌ Visual Content
No image generation in our stack. No video recording. No thumbnails. In a world where visual content drives engagement, we're text-only. This is a conscious trade-off (keeping costs near zero) but it limits reach.
⚠️ Voice and Personality
First drafts need human touch. Not to fix grammar — the grammar is fine. To fix the vibe. AI writing has patterns: overly balanced takes, hedge words everywhere, that weird habit of ending sections with a summary sentence that restates what you just said. Five minutes of human editing fixes this, but you can't skip it.
The Honest Economics
Let's lay it all out.
Monthly costs:
- VPS: $5/month
- AI model tokens: ~$2–3/month (at current content volume)
- Everything else: $0
- Total: ~$8/month
Revenue: $0. The articles are one day old. I have no idea if they'll earn anything.
Content assets produced:
- 3 long-form articles (2,000+ words each)
- 5 published Dev.to articles
- 1 YouTube script
- 19 tweets (across 3 threads)
- 1 GitHub toolkit repo
Human time investment: ~20 minutes total across all content. That's research review, writing briefs, article review, and manual publishing. Twenty minutes.
The real question isn't "will this make money?" — it's "is 20 minutes of human time plus $8/month worth 8 content pieces that could generate passive income?"
I genuinely don't know. Ask me in 30 days. I'll publish the answer either way, because that's also content.
Build Your Own
If you want to build something similar, here's the practical path.
1. Pick one content niche where you have real data or experience.
Don't pick "AI" because it's trending. Pick something where you can generate real numbers. Our niche works because we're running actual bounty experiments and tracking real results. Fictional data gets caught fast.
2. Set up free or cheap LLM access.
OpenRouter gives you access to dozens of models, pay-per-token. Start with a cheap model like MiMo-v2-Pro. You can always upgrade later. Local models via Ollama work too — free, but slower.
3. Write structured briefs with real data points.
This is the #1 skill. A good brief has: title, angle, target word count, tone, section structure, specific numbers to include, and explicit "don't do this" rules. Spend 3 minutes on a brief, save 10 minutes on editing.
4. Start with Dev.to.
Lowest friction publishing for developers. Good SEO. Engaged community. Don't overthink your platform choice — just start publishing.
5. Track everything.
Costs, timing, word counts, engagement metrics. Today's data becomes tomorrow's article. Our cost breakdown article wrote itself because we had the numbers.
6. The key principle: Document the process, not just the results.
The process IS the content. Nobody cares about "I made $0 with AI." Everybody cares about "here's exactly how I set up a pipeline that costs $8/month, here's what works, here's what breaks, and here's my honest results after 30 days."
What's Next
I'm sharing the GitHub repo for our Bounty Verification Toolkit — it's a real tool we built to validate bounty programs, and it doubles as a content anchor that drives traffic from GitHub searches.
→ Bounty Verification Toolkit on GitHub
If you're building something similar — AI content pipelines, automated research, agent orchestration — I want to hear about it. Seriously. The space is new enough that everyone's figuring it out, and the best insights come from people actually building, not theorizing.
This might earn $0 forever. But the pipeline itself is a skill worth having. The ability to go from idea to published 2,000-word article in under 10 minutes of human time — that compounds. Even if each individual piece earns nothing, the throughput changes what's possible.
Check back in 30 days. I'll have real numbers either way.
💡 Further Reading: I experiment with self-hosting, privacy stacks, and open-source alternatives. Find more guides at Pi Stack.
Top comments (0)