A scheduled agent watches competitor pages and writes a daily brief

The task

The company tracks 11 competitors across pricing pages, changelogs, blog posts, social accounts, and job boards. According to the author, one person spent Tuesday mornings clicking through more than 40 bookmarks, capturing changes, and writing a summary that was read days later. The company also paid for page-monitoring tools, a few automations, and mention alerts.

What they did

The author replaced the manual round with a Python script that runs on a daily schedule at 6 AM UTC on a hosted virtual machine. Each run has three steps. It fetches a list of competitor pages, compares them with the stored earlier version, and records the differences. It uses the Exa search API to find mentions, blog posts, launches, and GitHub activity from the last 24 hours that page checks would miss. A language model then reads the differences and search results and writes a brief grouped by competitor: what changed, what was published, what was said on social media, and an overall assessment with a confidence level. The first version flagged cookie banner changes as major updates and missed a pricing change because it watched the wrong page. According to the author, about a week of adjusting prompts and monitored pages made it dependable. The author also wrote a context document for each competitor so the assessments are less generic.

What the source reports

According to the author, research time fell from about 4 hours per week to about 15 minutes per day of skimming the brief. According to the author, the running cost is about $2 per day in API calls and compute, against $180 per month for the earlier tools. The author says the agent has run daily for 3 months. According to the author, it caught a quiet scheduling feature added to a competitor's documentation with no announcement, and the brief became the most-read document on the team.

Where the approval sits

The brief is internal reading material. The agent observes public pages and writes an assessment, and the source describes no message sent to competitors or customers. It also describes no review step: it does not say who checks the assessments before the sales and product teams use them. Clients of this library keep a person reading and deciding before anyone acts on a brief.

Worth knowing

The source is a blog post from a company that sells hosting for scheduled agents and ends with a pitch for it, so the savings are the author's own estimates. The author says page scraping is fragile: about once a month something breaks and a selector or scraping approach needs updating. The quality of the assessments depends on the context documents the author writes.

Prompt for your agent

More in this category

• Research and briefings

  A Claude Cowork skill builds one morning brief from calendar, chat, and tasks

  A team built a Claude Cowork skill that gathers the day's calendar, recent messages, and open tasks into one brief each morning. The brief is a summary: a person reads it, and the post describes no step where anything is sent.

  Read the entry
• Research and briefings

  A scheduled AI briefing replaces a manual morning round of inbox and calendar

  A technology executive has Claude Cowork prepare a briefing of urgent email, the day's calendar, and a short action list every morning. The author reads it, and the post shows no step where anything is sent.

  Read the entry