Ramp saved more than 30,000 manual hours and grew scale 5x after replacing human portal logins with Browserbase browser agents, Sahnan said on AI Engineer while demoing a verified Chromium fork built to survive antibot checks.
Before that shift Ramp and Bill relied on people to log into Delta, vendor portals and Secretary of State sites, a manual step that Bird argued could not handle thousands of concurrent sessions from a data center.
Bird said the team forked Chromium to remove the webdriver flag, rename headless Chrome to Chrome and keep the browser process alive between tasks instead of killing it, changes he claimed reduce bot detection that normally flags fresh fingerprints, empty cookies and data center IPs.
For control the pair showed Stagehand, an open source SDK that sits as a Chrome extension next to the browser and shrinks a 122,000 token raw HTML payload to about 4,000 tokens through an accessibility tree before asking a model where to click, according to their live comparison.
Stripe now uses the same pattern to verify merchants across state portals and Bill uses one Stagehand agent to handle thousands of different checkout flows, though the talk shared no error rates, capture rates or cost per verified filing and what would change the read is a published per portal success rate.
Parallel, which now powers search inside Browserbase, describes the product as the complete platform to build and deploy agents that browse the web like humans, and the open question is whether the 122,000 to 4,000 token cut and verified browser hold up when portals change their DOM and antibot rules every two weeks.
