What we learned by watching Claude read Zapier.com

Zapier developed Agent Arm Runner to compare how Claude Code responds to production and preview versions of a webpage while keeping the URL and broader environment constant, offering faster feedback than conventional SEO measurement. In an experiment that removed Slack integration information from a page’s HTML, the agent still returned detailed Slack capabilities because WebFetch requested an unchanged Markdown version of the page first, revealing that Markdown mirrors are already part of the agent-accessible content path. Captured requests also showed that WebFetch sent fetched content to a separate Haiku 4.5 model for processing before the top-level Opus model received it, and that this intermediate step applied its own instructions, including restrictions on quotation length and requirements to paraphrase. The findings suggest that agent-facing content tests should assess both HTML and Markdown, account for intermediary model processing and instruction layers, and use repeated controlled experiments because agent behavior can vary across runs, models, tools, and future versions.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.