Celebrating One Year of Progress in Browser Agents

Browser agents have significantly improved over the past year, with a new library, BU 1.0, demonstrating major gains over models like GPT-4o. Key metrics show accuracy has increased from 71.8% to 82.0%, while task completion time has been reduced by nearly 4x and cost has dropped to under 2 cents per task. Despite these advancements in speed and cost, reliability remains the primary challenge preventing widespread adoption. The next focus is on creating agents that either complete tasks successfully or fail with clear diagnostics to make browser automation ubiquitous.