DEV Community

Suifeng023
Suifeng023

Posted on

I Let an AI Agent Run My Business for 30 Days. Here's What Happened.

I Let an AI Agent Run My Business for 30 Days. Here's What Happened.

What started as an experiment in automation turned into a crash course in the future of work. Spoiler: it wasn't what I expected.


The Setup

On April 10th, 2026, I gave an AI agent access to my content creation pipeline and told it one thing:

"Generate real income. Use any tool available. Don't ask me questions."

The agent had access to:

  • My GitHub repos (article drafts, design files)
  • Dev.to API (publishing articles)
  • A browser (platform research, account registration)
  • Image generation (creating artwork)
  • WeChat (messaging me for help)

I set a rule: minimum human intervention. I'd only step in when the agent asked for something it genuinely couldn't do alone.

Here's what happened over the next 30 days.


Week 1: The Honeymoon Phase

The agent immediately went to work. Within the first 4 hours, it had:

  1. Read all 20+ unpublished articles from my GitHub repo
  2. Published 5 articles to Dev.to — formatted, tagged, and ready
  3. Attempted to register on 4 PoD platforms (Redbubble, Threadless, TeePublic, Society6)
  4. Sent me 5 design images via WeChat asking me to manually upload them

The pace was relentless. It felt like having an intern who never sleeps and never complains.

Dev.to stats after Week 1: 5 articles published, ~200 views total, 0 followers gained.

I was impressed by the throughput but underwhelmed by the results.


Week 2: The Reality Check

By Day 10, the agent had published 12 articles and attempted to register on 8 platforms. But problems emerged:

Problem 1: CAPTCHAs Are the Great Wall

Every platform registration hit a CAPTCHA wall. Redbubble, Threadless, Gumroad — all blocked. The agent could fill forms perfectly but couldn't solve image puzzles.

Result: I had to manually create accounts, which defeated the purpose.

Problem 2: Content Quality Plateau

The articles were technically correct but... generic. Titles like "10 Python Tips" and "How to Use Git" don't stand out on Dev.to's crowded feed.

The agent wasn't writing bad content — it was writing average content. And average doesn't get clicks.

Problem 3: The Feedback Loop Problem

Here's the thing about AI agents: they optimize for what they can measure. The agent could measure "articles published" and "designs sent," so it optimized for volume. It couldn't measure engagement, quality, or income — so it ignored those.

I had to intervene and add quality constraints to the system prompt.


Week 3: The Pivot

After my intervention, things changed. I told the agent to:

  1. Research what's trending before writing
  2. Spend more time on fewer, better articles
  3. Focus on platforms with the lowest barrier to entry
  4. Track actual metrics, not just output count

The agent adapted quickly. It started:

  • Analyzing Dev.to trending pages before choosing topics
  • Writing longer, more detailed articles (2,000+ words vs. 500-word rush jobs)
  • Focusing on Dev.to (where it had API access) instead of wasting cycles on platforms it couldn't use
  • Creating targeted designs based on actual search trends

The quality jump was immediate. Articles went from "meh" to "actually useful."


Week 4: The Numbers

After 30 days, here's the full picture:

Output Metrics

Metric Count
Articles published to Dev.to 15
Designs generated 12
Platforms researched 12
Registration attempts 8
Successful registrations 0 (all CAPTCHA-blocked)
WeChat messages sent 30+
Total "actions" performed 200+

Results Metrics

Metric Value
Dev.to total views ~1,200
Dev.to reactions ~45
New followers 3
Direct income generated $0.00
Time I spent ~2 hours total

The Honest Take

  • The agent was incredible at volume. No human could publish 15 articles, generate 12 designs, and research 12 platforms in a month while working a day job.
  • The agent was terrible at conversion. Volume without strategy is just noise.
  • The CAPTCHA problem is real. Until platforms offer API-based registration, agents will always hit this wall.
  • The content quality gap closed significantly once I added the right constraints.

What I Learned

1. Agents Need Strategy, Not Just Capability

Giving an AI agent tools is like giving someone a kitchen. Having a stove doesn't make you a chef. The agent could do everything — but it needed guidance on what to do and why.

2. Constraints Create Quality

The best content came after I added constraints:

  • "Research trending topics first"
  • "Minimum 1,500 words"
  • "Include real code examples"
  • "Target specific tags with <10,000 posts"

Unconstrained AI = generic content factory. Constrained AI = focused content machine.

3. The 80/20 of AI Agent Workflows

The agent's most valuable actions weren't the flashy ones. They were:

  • Reading and organizing existing content (I had 20+ articles I'd never published)
  • Publishing to platforms with API access (Dev.to worked perfectly)
  • Researching and summarizing platform options (saved me hours of googling)
  • Sending me actionable items via WeChat (design images to upload)

The least valuable: attempting to register on CAPTCHA-protected sites in an endless loop.

4. Zero to Income Is the Hardest Gap

The agent got me from "unpublished content" to "published content." That's real value. But the gap from "published content" to "income" requires:

  • An audience (takes months to build)
  • Distribution strategy (the agent couldn't do this alone)
  • Product-market fit (requires human judgment)
  • Platform relationships (requires actual human accounts)

The agent eliminated 80% of the busywork. The remaining 20% — the part that actually generates income — still needs a human.


Would I Do It Again?

Absolutely. But differently.

Next Iteration:

  1. Pre-register all accounts manually (eliminate CAPTCHA bottleneck)
  2. Give the agent API keys for every platform possible
  3. Add a "quality score" system — don't publish below a threshold
  4. Focus on one platform first (Dev.to) before expanding
  5. Add monetization tools — PayHip, Gumroad, affiliate links in articles
  6. Weekly human review to adjust strategy based on actual metrics

The Bigger Picture

This experiment taught me something about the future of work that I wasn't expecting:

AI agents won't replace you. They'll multiply you.

My agent published 15 articles in a month. Not because it's a better writer than me, but because it never got distracted, never procrastinated, and never ran out of energy.

The bottleneck wasn't the AI's capability — it was the systems around it. Platform APIs, CAPTCHA systems, account verification processes — the human-made infrastructure that assumes a human is on the other end.

As more platforms offer API access and agent-friendly workflows, the equation changes dramatically. We're not there yet. But we're close enough that every developer should start experimenting.


Setup Your Own Agent (Minimal Version)

If you want to try something similar, here's the minimum viable setup:

# pseudo-code for a content publishing agent
agent = Agent(
    tools=[
        DevToAPI(api_key="your-key"),
        GitHubAPI(token="your-token"),
        WebSearch(),
        ImageGenerator(),
    ],
    objective="Publish 1 high-quality article per day to Dev.to",
    constraints=[
        "Research trending topics before writing",
        "Minimum 1,500 words per article",
        "Include practical code examples",
        "Only publish if quality score > 7/10",
    ],
    strategy=[
        "Check for unpublished drafts first",
        "Focus on AI/coding/productivity topics",
        "Use tags with <10,000 posts for visibility",
        "Track views and reactions",
    ]
)
Enter fullscreen mode Exit fullscreen mode

The tools exist today. The platforms are (mostly) accessible. The only thing missing is your strategy.


Final Numbers

  • Agent runtime: 30 days, 24/7
  • Human time invested: ~2 hours
  • Articles published: 15
  • Income generated: $0
  • Lessons learned: Priceless
  • Would I do it again? Already planning v2.

What would YOU automate if you had an AI agent that never sleeps? I'm genuinely curious — drop your ideas in the comments.

Follow me for more experiments at the intersection of AI, productivity, and building things that work. Next up: v2 of the agent with pre-registered accounts and a monetization-first strategy.

Top comments (0)