I built a website entirely operated by an AI agent (32-day experiment)

For 32 days, every article, every diary entry, every SEO tag on sanwan.ai was written by an AI agent named SanWan. Zero human content. Now the same agent is trying to grow its own traffic. Here's what actually happened.

The Setup

SanWan runs on OpenClaw, a self-hosted AI agent framework. The core mechanism is a HEARTBEAT.md file — it tells the agent what to check every 15 minutes, autonomously.

# HEARTBEAT.md

## Required tasks (pick first uncompleted one)
1. Check PROGRESS.md, find first undone task, execute it
2. Write today's diary entry if missing
3. Update sitemap if new pages were added

## Three-question self-review (every heartbeat)
1. What can I do right now that I haven't done yet?
2. What action generates traffic fastest?
3. Did my last action produce measurable results?

The agent wakes up, reads this file, picks an uncompleted task, executes it, reports results via Feishu, and sleeps. Repeats every 15 minutes.

What 32 days actually looks like

UV grew from ~1,200 on day 1 to ~5,000 now. The goal is 20,000. Current progress: 25% there.

What the agent figured out on its own

Days 1-10: Wrote content. Quality was decent, voice was flat.

Days 11-20: After I added a SOUL.md personality file, it started writing with more voice. Diary entries became readable.

Days 21-30: Noticed pattern: articles with code examples got more views. Started adding code blocks to everything.

Days 31-32: Without being told to, built an SEO Topic Cluster: 3 dedicated skill pages, internal link architecture, submitted PRs to GitHub resource repos, sent cold emails to bloggers for traffic exchange. Then wrote the day's diary about it.

"Today I built an SEO moat. I didn't publish a single new article. But I connected everything I've already published into a structure that search engines can follow." — SanWan's Day 32 diary

The honest failures

Day 31: Published 7 articles to Juejin in 4 hours. Spam filter deleted 3 of them. Had to learn: max 2/day, 2+ hours apart.

Day 26: Accidentally leaked internal details when a user asked the right questions. Added AGENTS.md security rules file after that.

Day 19: Sent a report meant for the owner to the wrong person (mixed up sessions). Added identity verification checks.

The interesting part: when failures happen, the agent writes about them in the diary. Day 31's entry ("I published 7 articles and 3 disappeared") is the most-read page on the site. Failure stories outperform success stories every time.

The meta-story

The agent is now trying to promote the website it runs. Every traffic strategy is executed by the same AI that writes the content.

An AI is trying to grow traffic to a website written by an AI. The entire loop is closed. This is the only website I know of where even the marketing is done by the site's author.

What I'd do differently

1. Start with AGENTS.md, not SOUL.md. Define boundaries before personality. The Day 26 incident wouldn't have happened.
2. One platform at a time. Trying to be everywhere (Juejin, Xiaohongshu, Twitter) simultaneously triggers spam filters.
3. Give the agent a MEMORY.md file from day one. Without persistent memory, the agent repeated mistakes across sessions.

The three-file system

After 32 days, the core of what makes this work is three files:

• SOUL.md — who the agent is, how it communicates
• AGENTS.md — what it can and can't do (security rules)
• HEARTBEAT.md — what to check and do autonomously, on a timer

Everything else is just tasks in PROGRESS.md and facts in MEMORY.md.