Summary
Corey Gallon of Rexmore presents a technical talk at AI Engineer on making AI agents interact with websites the same way humans do — an increasingly necessary capability as anti-bot systems from Cloudflare, Google, and individual sites become more sophisticated. The central premise is that using the Chrome DevTools Protocol (CDP) to drive a real Chrome browser means that agent clicks and keystrokes travel the same internal path as genuine human input, making them indistinguishable at the browser level.
Gallon introduces his open-source Chrome Agent tool and frames the implementation approach as three components: a CLI (not MCP) for driving the browser, the CDP itself, and a “loop on a ladder” sense-act-verify cycle. The CLI versus MCP comparison is backed by data from an Arise AI study: both achieve roughly 83% task success, but CLI completes the same task in 7 turns and under one minute versus 71 turns and 8 minutes for MCP, with Anthropic reporting CLI can be up to 75 times cheaper in token cost. He also notes that OpenAI flagged his own account for “cyber abuse” while he was preparing the demo, underscoring how real the detection arms race has become.
The “meatag ladder” framework he presents escalates trust levels of browser interactions from synthetic JavaScript clicks (rejected by sites like Amazon’s add-to-cart button as “untrusted”) through CDP input events (stamped trusted by Chrome’s internals) up to full human input path simulation. Practical examples include bulk personalized email sending via Outlook web UI and e-commerce automation, with the web UI framed as a “permissionless API” for corporate environments where REST API access requires admin approval.
📺 Source: AI Engineer · Published August 14, 2026
🏷️ Format: Tutorial Demo
