Hey folks, I've been playing with Relevance AI's agent-building studio for a few weeks now, mostly to automate some of my data source research. I finally built something pretty neat and wanted to share.
I made an agent that takes a list of website URLs, visits each one, captures a full-page screenshot, and then writes a concise summary of the page content. It's been super handy for quickly assessing potential data sources or competitor tools without manually opening twenty tabs. The visual context from the screenshot combined with the summary gives me a much better feel than a plain text scrape.
Here's a peek at the core agent logic I set up. It uses the `web-scraper` tool for the content and a custom action for the screenshot, chained together with a prompt for the summary.
```yaml
agent:
name: Website Screenshot & Summarizer
tasks:
- task: navigate_and_capture
tool: web-scraper
parameters:
url: {{input_url}}
output: raw_html
- task: take_screenshot
tool: custom_screenshot_action
depends_on: navigate_and_capture
parameters:
html: {{raw_html}}
output: screenshot_url
- task: generate_summary
tool: llm
depends_on: navigate_and_capture
parameters:
prompt: "Summarize the key content from this webpage in 3 bullet points: {{raw_html}}"
output: text_summary
```
The studio made it straightforward to wire this together. The biggest "aha" was getting the dependency chain right so the summary uses the scraped HTML. I've run it on a mix of blog posts and SaaS landing pages, and the results are solid for a first pass.
Has anyone else built agents for web research or content digestion? I'm curious about different approaches to handling pagination or dynamic content. Also, wondering how you all are managing the output—I'm piping the summaries and screenshot links into a BigQuery table for now.
ship it
ship it
That's a clever combination of tools. I've built similar workflows for monitoring client dashboards and marketing sites, and the visual context from a screenshot is indeed invaluable. It can immediately reveal layout issues or dominant visual elements that pure text analysis would miss.
One point you might consider for production use is handling dynamic content. The `web-scraper` tool fetching `raw_html` before the screenshot means you're capturing the page state at that exact moment. If the page has client-side rendering or interactive elements that load after the initial HTML, your screenshot might not match the intended summary. A possible alternative is to drive both the scrape and screenshot from a headless browser instance in a single step, ensuring they work from the same DOM state.
Also, have you thought about error handling for inaccessible sites? Building in a timeout and retry logic for the `navigate_and_capture` task can prevent one failed URL from stalling the entire list.
null