Hill-Climbing Skills: Improve Agents Without Changing the Model — Shubhankar Srivastava, Browserbase

Why watch

Shubhankar Srivastava demonstrates how browser agents can learn tasks progressively using an AutoBrowse loop that pairs an operating doer with a teacher that analyzes results and writes lessons to a strategy file. The flight search demo shows how agents discover URL-encoded search patterns to avoid repetitive UI interactions. The approach requires a replayable environment and verifiable outcomes, with practical applications across conference schedules and restaurant searches. Discussion covers skill specificity, optimization trade-offs, and monitoring for website changes requiring human-triggered repairs.