On November 20, 2026, the open-source browser agent Browser Use published a retrospective explaining why version 1.0, released on November 10, dared to delete code: 24 browser actions were collapsed into a single python tool, the codebase lost 136,905 lines and gained only 29,601, and 178 Python files shrank to 43 — while the same models scored 14 to 33 points higher on BU Bench v2.1 and cost 2.7 to 5.2 times less.
Fewer tools, more initiative from the model
Version 1.0 gives the model one main tool: a persistent Python environment with pre-imported browser helpers and raw Chrome DevTools Protocol (CDP) access, plus two ways to finish. For two years the team had followed a "stuck, add an action" approach, and by June 2025 the tool count had grown to 29; every action occupied prompt space, and anything outside the menu was unreachable for the model. Once it could write code instead, the model made 12,976 code calls across 199 benchmark traces, using about 47 CDP commands: when a page hung, it blocked the offending script, read the page's own JavaScript, found the data endpoint behind it, and pulled back 4,720 rows in three and a half minutes.
Deleting code is not unconditionally right. The author warns that 1.0 also changed default settings, so the score gains cannot be credited to the tool change alone; and captchas remain a place where code loses — in a small test, a prose prompt failed all three hCaptcha attempts, while the dedicated solve_captcha tool passed six out of six. The team's rule is therefore direct: if the model can already reach an underlying capability, write it as prose and let it use it; only build a tool for what it cannot reach.
Getting started with 1.0 is easy: install the package, add an API key, and three lines of code are enough to run it. The lesson reaches beyond browser agents: the goal of tool design is not to give the model a longer menu, but primitives that reach deep enough. Measured against this site's earlier review, billing and login walls were the first pitfalls users hit with Browser Use — whether 1.0 shrinks them will depend on real-world feedback. Is Browser Use worth using? Letting an agent click through web pages itself — the pitfalls are billing and login walls is worth reading alongside.