browser-use 1.0 Replaces 24 Agent Tools With One Python REPL, Scores Jump 14-33 Points
lxfater · x · 2026-11-21
browser-use 1.0 throws away the 24 browser actions of 0.x and gives the model one main tool, python: a persistent REPL with raw Chrome DevTools Protocol access pre-imported. The release deleted 137K lines of code.
Why: a big tool menu burns tokens every turn, and the agent deadlocks whenever a page falls outside the preset actions. With code instead of a menu, the same three models scored 14–33 points higher on BU Bench v2.1, cost 2.7–5.2x less per task, and ran 2–3x faster (the team admits defaults also changed, so the tool change can't take all the credit). The agent started unblocking hung pages, killing stuck scripts, and calling private APIs nobody had built actions for.
Not everything should be a REPL: on CAPTCHA tasks prompts failed 3/3 while a dedicated tool passed 6/6. Rule of thumb: if the model can reach it with code, let it code; otherwise build a tool.
Caveat: the REPL is not a sandbox — it runs with your local privileges, so wrap untrusted tasks in Docker or a VM.
More from coding & agent
- AI gateways are becoming the key control layer in production AI stacks: 8 capabilities — goyalshaliniuk · 2026-11-21
- Mnemos 2.0: agent memory designed to preserve what matters for the next session — RileyRalmuto · 2026-11-21
- InferenceBench ranks agents on LLM serving speedup: Claude Opus 5.5 tops at 12.08x, agents lose to simple hyperparameter search — maksym_andr · 2026-11-21
- 4 architecture patterns to tame multi-agent system reliability, from a 27-year reliability engineer — bibryam · 2026-11-21
- AutoCut + FrameAtlas: open-source local-first video editing driven by Claude Code via MCP — moheshmohan · 2026-11-21
- InfoQ: What an agent harness is, and two ways to build one — bibryam · 2026-11-21