The AI agent space is full of demos. Tools that publish a tweet autonomously, or summarize a document, or answer a question in a Slack thread. Impressive in a slide deck. Less impressive when you're trying to actually run a business.
OpenClaw is different — and I've been using it to run six websites for the past few weeks.
Here's what it actually does.
The Setup
Six sites: an AI news publication, a free course platform, a fitness association, a portfolio, a consulting blog, and a free library. Different audiences, different content cadences, different monetization strategies. Previously, keeping all of them updated required constant attention. Now, most of it runs without me.
The core of OpenClaw is a heartbeat system — a HEARTBEAT.md file that defines what the agent should check and do on every scheduled poll. Think of it as a living task list that the agent reads, acts on, and updates. Combine that with cron jobs (scheduled tasks that fire at specific times) and subagents (isolated agents spawned for specific jobs), and you have a system that can handle a surprisingly complex workload.
What Runs Every Day
Content at 11:00, 16:00, and 20:00 UTC: The agent fetches fresh AI news from sources like The Verge, identifies stories not yet covered, writes 2-3 full articles with headlines, body copy, FAQs, and cover images (via Gemini's image generation API), runs the test suite, and pushes to GitHub. Every day. Without me asking.
Social media drafts at 12:00-15:00 UTC: Four emails land in my inbox — a LinkedIn post for the news site, three tweet drafts, a LinkedIn post for the course platform, and a tweet for that too. I copy-paste what I want to use. The AI handles the writing; I handle the judgment call on what to actually post.
Monday morning strategy brief: Every Monday at 11:00 AM, an email arrives with the week's traffic data across all sites, the top easy-win keywords (ranking 8-20, within reach of page one), content priorities with specific post titles, SEO issues to fix, a feature suggestion with a ready-to-paste Claude Code prompt, and three questions that only I can answer. It's a proper business review, generated entirely from real GSC and analytics data.
Retention emails: Users who haven't completed a lesson in a week get a re-engagement email. Users whose learning streak is about to expire get a reminder at 20:00. Both automated.
What It Actually Built
Beyond content, the agent has shipped real code:
- A testimonials section for a client's detailing website, populated with real Google reviews
- Individual service pages with SEO content and FAQ schemas
- A streak celebration toast that fires after lesson completion on the course platform
- Two API routes for automated email campaigns
- 301 redirects after consolidating cannibalizing blog posts (identified via GSC data analysis)
Each of these came from a conversation: "build this, make a PR." The agent (using Claude Opus for complex features) reads the codebase, writes the code, runs the tests, and opens a pull request. I review and merge.
What Doesn't Work Yet
Telegram exec approvals have been a recurring headache — the agent can't run arbitrary shell commands from the chat interface without configuration. The workaround is spawning subagents that have full permissions, but it adds friction.
The agent also can't post to social media directly — no browser automation available. The workflow is drafts-by-email, manual posting. Good enough for now, but a gap.
And like any AI system, it occasionally generates duplicate content (we had four pairs of duplicate articles before adding a de-duplication check) or misses context that a human would catch. The answer isn't less automation — it's better guardrails and regular audits.
The Honest Verdict
OpenClaw isn't magic. It's infrastructure. You still need to define what you want, set up the workflows, review what it produces, and make the judgment calls it can't. The Monday strategy brief is useful because I read it and act on it — not because it acts autonomously.
But for someone running multiple content properties without a team, the leverage is real. Articles published daily without lifting a finger. SEO analysis that would take hours done in minutes. Code shipped without context-switching into a development environment.
The open-source AI agent era is here. The question isn't whether to use it — it's how to configure it for your specific situation.
OpenClaw is available at openclaw.ai. The community Discord is active and the documentation is solid for a project of this stage.
