New Tools for Browser Automation: BrowserSkill and Browser Use's Bitter Lesson

Browser Use co-founder Greg PR demonstrated that as models improve, browser automation harnesses should be stripped down to raw CDP, allowing models to write code directly. Separately, Tencent open-sourced BrowserSkill, a CLI+browser extension that lets AI agents use the user's real browser with login state, without interruption, via a secure four-layer architecture.
Key insights
Browser Use's Evolution: From Predefined Actions to Raw CDP
Browser Use evolved over two years: initially defining state and action spaces manually, then introducing JavaScript execution (replacing 12 browser tools with one), then removing predefined state extraction by letting models directly access CDP. The key insight: as model coding ability crosses a threshold, human-designed intermediate representations become liabilities. The final principle: expose the simplest raw interface (CDP) and let the model choose observation and actions.
Tencent's BrowserSkill: Using the User's Real Browser with Login State
BrowserSkill allows any AI agent that can run shell commands to use the user's actual browser with full login state, without interrupting the user. It operates via a four-layer architecture: Agent → bsk CLI → bsk daemon → browser extension → Agent Window. The agent never directly touches the browser. It supports remote mode (agent on server, browser on local machine) with TLS-authenticated WSS.
BrowserSkill's Security Model: Sandbox and Credential Protection
All write operations are sandboxed to the Agent Window unless explicitly borrowed with user confirmation. Automation switches (borrow confirmation, help request) are stored in the browser and cannot be overridden by CLI. File transfers split authority: CLI handles file paths, daemon manages storage, extension never sees paths. Credentials are never extracted; cookies remain in the browser profile. Remote credentials store only hashes. 30-day audit logs are maintained.
Agent Engineering Discipline in BrowserSkill
The SKILL.md includes defensive rules: navigation invalidates references, so re-observe before interacting; if borrowing is denied or help canceled, stop; never bypass security; screenshot evidence and page assertions are verified separately; Canvas interactions use single-use capture_id with 2-minute expiry. Installation is agent-first: user says 'install browser-skill' and the agent completes setup, skill injection, and verification.
Evaluation Framework and Engineering Maturity
BrowserSkill includes a thorough evaluation framework with local deterministic pages, agent-neutral prompts, seeded DOM variants, and honest verification (unverified results are not silently counted). 28 browser operations covered across five test suites. The project has stable iteration (0.3.0 in ~3 months), cross-platform releases, and CI automation. Codebase: ~3.5M lines TypeScript, ~1.7M lines Rust, with i18n in English, Chinese, Korean.