How to Build Financial Agents That Work Across the Web

Jay Sahnan, Browserbase, Arik Bird, Browserbase19:38 · Oct 2026 · 10K views
Thumbnail for How to Build Financial Agents That Work Across the Web Watch on YouTube
TL;DR
  1. 1

    Browser agents give financial companies access to data behind logins, anti-bot systems, and sites without APIs.

  2. 2

    Browserbase runs a modified Chromium in the cloud with persistent browser processes and changes that make data-center browsers look more like ordinary browsers.

  3. 3

    Stagehand combines browser control with natural-language actions, DOM reduction, and structured extraction so one agent can work across changing websites.

Summary

Jay Sahnan and Arik Bird explain why financial automation often needs a browser rather than an API. Ramp uses browser agents to log into vendor portals, fetch receipts, and match them to card transactions. Bill needs one agent to handle thousands of portals with different layouts and payment flows. Stripe uses browser automation for KYC checks through Secretary of State sites. The talk also explains Browserbase's infrastructure. Its Chromium fork removes automation signals, aligns the browser identity, and keeps browser processes alive between tasks. Stagehand sits above the browser and supports deterministic Playwright-style commands alongside natural-language actions. Its observe, act, and extract primitives help agents find controls, perform tasks, and turn messy web pages into structured data. Demos show 100 concurrent browsers and the same research workflow running across 20 companies.

Key ideas
00:53

A browser is the agent's tool for sites without APIs

Sahnan and Bird define a browser agent as an agent that can access, read, and act on the web through a browser. They use examples such as booking a flight, checking competitor pricing, and planning a date night. The browser becomes especially important for sites behind logins, anti-bot protection, or portals that expose no API or MCP. Their practical claim is that any agent expected to do work on the web should have browser access.

02:11

Ramp replaced manual receipt collection with browser agents

Ramp first handled receipt collection by having people log into difficult vendor portals. That approach worked for a smaller company but became a scaling limit as the number of portals grew. After a card transaction, Ramp can start a browser agent that visits the relevant site, logs in, captures the receipt, and matches it to the transaction. Sahnan says Ramp now deploys thousands of browser agents, has saved more than 30,000 manual hours, and has increased its scale fivefold.

03:47

Data-center browsers expose signals that ordinary browsers do not

A laptop browser carries a persistent identity made up of its browser version, operating system, fonts, screen size, time zone, cookies, login state, and network reputation. A new headless browser in an EC2 instance instead gets a fresh fingerprint and cookies, may require another login, and reveals a data-center IP. Proxies can change the apparent location or reduce some CAPTCHA prompts, but Sahnan says they do not fix the underlying Chromium signals. This is why simply moving laptop-style Chromium into the cloud does not provide reliable scale.

06:03

Browserbase forked Chromium to keep cloud browsers consistent

Browserbase forked Chromium because its browser engine changes every two weeks and because the standard binary exposes automation and short-lived process behavior. The fork removes the automation flag, presents the identity as Chrome rather than headless Chrome, and keeps Chromium running after a task while closing its windows. Sahnan says these changes reach deep into Chromium and must be maintained as the upstream browser changes. The talk presents this modified browser as the base for reliable cloud automation.

07:49

Concurrent hosted browsers provide scale without manual control

The demo launches one hosted browser first and then repeats the same visit across 100 browsers at once. Attendees can open individual browser views through an iframe and interact with them by clicking, moving, and controlling the page. The speakers explain that live views are useful during development, while production workloads can run without someone watching every browser. This demonstrates the infrastructure needed for workflows such as visiting many sites concurrently.

10:11

Stagehand lets one agent adapt to many portal layouts

Bill processes payments across thousands of vendors, each with its own checkout flow, form, and changing interface. Handwritten Playwright scripts depend on selectors and can break when a button moves or a field changes from 'first name' to 'name'. Stagehand keeps deterministic browser commands but also supports natural-language actions such as filling a form or finding a red button. It sends the DOM and accessibility tree to a language model so the agent can identify the right control instead of relying on one fixed selector.

13:09

Stagehand reduces agent context and browser round trips

For one example, the speakers compare sending about 122,000 tokens of raw HTML with Stagehand's reduced accessibility tree, which uses about 4,000 tokens and provides direct selectors. They say this lowers token use and helps the agent respond faster. Browserbase also runs Stagehand next to the hosted browser as a Chrome extension. That placement avoids repeatedly sending page data back and forth between the browser and a remote controller, reducing action latency.

14:52

Stripe uses browser agents for KYC research

Stripe onboards many merchants and needs to verify that businesses are real and suitable to work with. Before browser agents, people visited Secretary of State portals and used the browser to collect the required information. The proposed replacement is a browser agent that performs the same KYC checks. Stagehand exposes three primitives for this kind of work: observe can find a control, act can click or fill it, and extract can convert web content into a Zod schema.

16:35

The same extraction workflow can run across many companies

The final demo extracts information from a company website, investor page, SEC filing, and New York Stock Exchange listing. It first runs the workflow for Chipotle, then runs the same code across 20 companies, including Chipotle, JPMorgan, and Goldman Sachs. The speakers say the browsers run in parallel and the live views may lag behind the work happening in the background. The extracted results appear in structured output, while the agent navigates the sites without a person watching each browser.

"You can extract anything into a Zod schema, which is great for engineers because we're taking really messy unstructured data on the web and converting it to a really easy-to-use data and readable and send it to an agent."Arik Bird16:16
Who should watch
  • You are building financial workflows that need information from portals with logins, anti-bot protection, or no API.
  • Your current automation relies on Playwright selectors and breaks when vendor websites change.
  • You need to run browser research or data collection across many sites at once while keeping structured results.