Browserbase’s Jay Sahnan and Arik Bird show how financial agents reach account portals and handle changing interfaces in this 20-minute AI Engineer talk.
Start with the browser infrastructure
- Receipts, vendor payments and business records often live behind logins or on sites without a suitable API.
- A fresh cloud browser can lose the cookies, consistent identity and network reputation that make a local session work.
- Browserbase maintains a Chromium fork and keeps browser processes alive between tasks to reduce that friction.
- A live demonstration launches 100 browsers; the speakers explicitly distinguish visiting websites from completing financial workflows.
Adapt to changing pages
- Stagehand, Browserbase’s open-source browser-control toolkit, combines ordinary scripted actions with natural-language instructions.
- It presents a compact page representation to the model, reducing irrelevant HTML and helping identify controls.
- The speakers show an example shrinking roughly 122,000 tokens of HTML to about 4,000 tokens of page structure.
- Hosting browser control near the browser reduces network round trips, according to the presenters.
Match each workflow to its job
- Ramp: collect receipts from account portals and match them to transactions. Browserbase reports over 30,000 manual hours saved.
- Bill: use an adaptable agent across different vendor payment interfaces, reducing the need for a separate script per portal.
- Stripe: collect business-verification evidence from public registries for know-your-customer checks.
- Research demo: gather company-site, investor-relations, SEC and exchange information, then expand the exercise to 20 companies.
Keep observation, action and evidence separate
observefinds possible page actions;actcarries out an instruction;extractreturns information in a specified data structure.- Reliable access and structured output still need permission checks, evidence validation and appropriate approval before financial actions.
- Caveat: customer savings and reliability figures are vendor claims. The demos supply no end-to-end payment error rate or independent accuracy audit.
“One agent, thousands of different portals.” — Jay Sahnan, 11:03