Browserbase's Jay Sahnan and Arik Bird appear in a YouTube talk on building browser agents for financial data locked behind logins and anti-bot walls. They present three case studies — Ramp receipt fetching, Bill vendor-portal payments, and Stripe KYC automation — plus Browserbase's forked Chromium and Stagehand primitives.
Derek Meegan, software engineer at Browserbase, gave a talk at AI Engineer World's Fair 2026 on moving browser agents from demos to production. He quantifies compounding step error: 99% per-step accuracy yields roughly 36% success over 100 steps. He then presents a production architecture for automating a health insurance portal that keeps the model off the deterministic path.
Derek Meegan, software engineer at Browserbase, gave a talk at AI Engineer World's Fair 2026 on moving browser agents from demos to production. He quantifies compounding step error: 99% per-step accuracy yields roughly 36% success over 100 steps. He then presents a production architecture for automating a health insurance portal that keeps the model off the deterministic path.
Miguel González Fernández (Browserbase) and Corby Rosset (Microsoft Research) presented work on verifiers for computer-use and web agents at AI Engineer World's Fair 2026. Their Universal Verifier replaces LLM-as-judge evals that overstated agent success (74% judge vs 38% verifier), cutting false positives from roughly half to near zero while matching human-human agreement. The talk accompanies a paper, blog post, and GitHub release including CUAVerifierBench.
Miguel González Fernández (Browserbase) and Corby Rosset (Microsoft Research) presented work on verifiers for computer-use and web agents at AI Engineer World's Fair 2026. Their Universal Verifier replaces LLM-as-judge evals that overstated agent success (74% judge vs 38% verifier), cutting false positives from roughly half to near zero while matching human-human agreement. The talk accompanies a paper, blog post, and GitHub release including CUAVerifierBench.
Stagehand v4, an open source SDK for browser agents, launched on Hacker News. It addresses limitations of Playwright for agent use, such as bloated context windows, iframe issues, and state sync problems. The SDK aims to give agents better internet access for knowledge work.
Stagehand v4, an open source SDK for browser agents, launched on Hacker News. It addresses limitations of Playwright for agent use, such as bloated context windows, iframe issues, and state sync problems. The SDK aims to give agents better internet access for knowledge work.
A developer reports confusion navigating Browserbase's product lines—Stagehand SDK, Browse CLI, and Browse.sh—which appear to serve different purposes but are documented interchangeably. The user spent 40 minutes on a Stagehand quickstart before realizing they likely needed the CLI, highlighting a discoverability problem.
A developer reports confusion navigating Browserbase's product lines—Stagehand SDK, Browse CLI, and Browse.sh—which appear to serve different purposes but are documented interchangeably. The user spent 40 minutes on a Stagehand quickstart before realizing they likely needed the CLI, highlighting a discoverability problem.
A YouTube talk argues that agentic AI systems lose roughly 70% of tokens processing CAPTCHA responses instead of real content, because HTTP 200 status codes mask failed scrapes. The problem goes unnoticed in most pipelines since status and response size appear normal, silently inflating cost and degrading agent output quality.
A YouTube talk argues that agentic AI systems lose roughly 70% of tokens processing CAPTCHA responses instead of real content, because HTTP 200 status codes mask failed scrapes. The problem goes unnoticed in most pipelines since status and response size appear normal, silently inflating cost and degrading agent output quality.