How to Build Financial Agents That Work Across the Web - Jay Sahnan & Arik Bird, Browserbase

하이라이트

• Financial institutions like Ramp, Bill, and Stripe rely on browser agents to access heavily protected websites and portals that lack APIs or model context protocols.

• Forking the Chromium binary to remove automation flags, align identity, and keep processes alive prevents data center browsers from being flagged by anti-bot systems.

• Stagehand shrinks raw HTML DOMs from 122,000 tokens down to 4,000 tokens using accessibility trees, significantly lowering LLM costs and execution latency.

• Stagehand's extract primitive converts messy unstructured web data directly into Zod schemas for reliable engineering workflows.

• Deploying browser agents saved Ramp over 30,000 manual data entry hours and increased operational scale by 5x.

타임라인

1. 00:00 - 02:13

   The Role of Browser Agents in Financial Operations

   Financial platforms require continuous interaction with complex web portals. Because many target sites lack public APIs, browser agents serve as the primary mechanism for reading data and executing actions across the web.

   • Financial institutions handle sensitive data stored behind login portals and anti-bot protections.
   • Browser agents enable automated access to websites that lack public APIs or direct integrations.
2. 02:13 - 09:43

   Scaling Browser Infrastructure in Data Centers

   Relying on human teams to fetch receipts and log into portals creates severe scaling bottlenecks as companies grow. Deploying standard headless browsers in data centers introduces detection risks because websites easily spot data center IPs and automation markers. Customizing the browser binary neutralizes these detection vectors, allowing organizations to run thousands of concurrent browser sessions reliably.

   • Running standard headless Chromium in data centers triggers anti-bot blocks due to distinctive data center fingerprints and missing cookies.
   • Forking the Chromium binary allows developers to remove automation flags like webdriver equals true and keep processes alive.
   • Ramp saved over 30,000 manual hours and achieved a 5x scale increase by replacing human data entry teams with browser agents.
3. 09:44 - 15:00

   Adaptive Web Navigation with the Stagehand SDK

   Processing payments across thousands of distinct vendor portals makes rigid automation scripts unmaintainable. Stagehand resolves this by sitting on top of Playwright, enabling agents to execute natural language commands like filling forms or locating specific elements. Positioning Stagehand as a Chrome extension directly beside the browser eliminates round-trip latency and cuts LLM token consumption dramatically.

   • Traditional Playwright scripts break frequently when target website layouts, button positions, or text fields change.
   • Stagehand combines deterministic code with natural language prompts to let agents interact dynamically with thousands of unique portals.
   • Compressing raw HTML DOMs into accessibility trees reduces token counts from 122,000 to 4,000 per request.
4. 15:01 - 19:27

   Multi-Source Data Extraction and KYC Verification

   Onboarding thousands of merchants requires verifying business legitimacy through multiple independent web sources. Using Stagehand primitives, developers can observe elements, perform actions, and extract structured data without manual oversight. Concurrent execution across dozens of corporate sites enables rapid data gathering for financial analysis and compliance verification.

   • Stripe automates merchant KYC verification by gathering company data across state portals, investor pages, and SEC filings.
   • Stagehand exposes three core primitives: observe, act, and extract.
   • The extract primitive maps unstructured web pages directly into Zod schemas for structured data processing.