At my day job I run product for a large classifieds marketplace. Marketing wanted to know what the internet says about the brand — and about competitors — without paying Brandwatch-tier money for it. So I built the monitor out of a computer-use agent.
The setup
The agent (built on OpenClaw, a computer-use framework I'd been testing for months) runs once a day:
- Walks through a list of news sites and social platforms — as a human would, in a real browser. No APIs, no scraping infrastructure, no per-platform integrations.
- Searches for mentions of our brand and key competitors.
- Classifies each mention: who is talking, in what context, what tone.
- Compiles a daily report with conclusions and recommendations for the marketing team.
The reason a computer-use agent fits this job: mention sources are messy and API-hostile. The agent doesn't care — if a page renders in a browser, it can read it. That's the one thing it does structurally better than a code-first assistant: it browses like a person.
The catch that justified the whole thing
Within weeks, the report flagged something no one was watching for: a cluster of user complaints about fraud experiences was forming on Reddit. Not on our support channels, not in app-store reviews — on a platform nobody on the team monitored.
That's an early signal of exactly the kind that turns into a PR crisis if it composts for a few months. Because it surfaced in a daily digest, the team could respond while it was still a conversation, not a narrative.
What I learned
Brand monitoring is a real, working use case for computer-use agents — not a demo. It produces value a marketing team acts on weekly.
Early signals live where you don't look. The value wasn't volume of mentions; it was one thread in the right place at the right time. Automated breadth beats manual depth for detection.
The cost structure changes the decision. Enterprise social-listening tools price this as a five-figure annual commitment. An agent running on infrastructure you already have prices it at approximately zero. At that price, "should we monitor?" stops being a question.
Agents like this are a PM tool, not just an engineering toy. No pipeline was built, no vendor onboarded — a product manager described what to check, and the checking got done every day since.
Originally discussed (in Russian) on my Telegram channel, where I share cases like this from work and side projects.