Why 99% Accurate Browser Agents Still Fail — Derek Meegan, Browserbase

Derek Meegan, a software engineer at Browserbase, gave this talk at the AI Engineer World's Fair 2026 in San Francisco. His starting point is a piece of arithmetic: a browser agent that is 99% accurate on each step succeeds only about 36% of the time across 100 steps. The talk is about the gap between a demo and thousands of unattended production runs.

After the intro, a chapter at 0:12 contrasts demo and production, and the agenda follows at 0:47. From 1:27 he explains how agents interact with the web, and at 2:32 he compares the three ways they see it: text, screenshots or code. Types of browser trajectories come up at 3:17.

The core argument starts at 4:32 with "no partial credit", then the 99%-to-36% problem at 5:22. At 6:12 he frames the economics as profit equals revenue minus cost: according to the description, cost builds up with every step while value only arrives when the task finishes. He then ranks performance first, with cost and maintainability as engineering problems after it (7:37).

Chapters at 8:12 and 8:42 define success. His recommendation is to measure success per transaction rather than per run, and the description notes that retries change the picture. Cost is broken down at 9:32, maintenance over time at 10:11, and at 10:46 he covers the risks of a web that fights back against automation.

The last third is a worked architecture for automating a health insurance portal, starting at 12:06. It adds a deterministic download tool (13:01), verification with OCR (13:31), authentication pulled out of the agent (13:56) and skills used as standard operating procedures (14:36). The final system appears at 15:26, and the talk ends at 16:21. The overall idea, per the description, is to take everything deterministic out of the model's hands.

The description links Browserbase and its open-source Stagehand framework on GitHub. The talk suits engineers building web automation who need numbers to argue for fewer model-driven steps.