Engineering Blog
Production practices for proxy reliability, anti-blocking, compliance, and cost optimization.
Agent A recorded its operation state. Agent B doesn't know. Agent A's session needs to migrate to another node, but cookies and localStorage are still on the original browser.
The Executor's core is a for loop: Planner runs every N steps to guide direction, Navigator executes one operation per step. Planner says done → stop. Navigator says done → Planner validates. Consecutive failures hit the limit → abandon.
Install, run, and bypass Cloudflare Turnstile with CloakBrowser in 3 lines of code.
Profile isolation works but is fiddly. Containerization is thorough but expensive. CDP Target is lightweight but fragile. Three approaches compared by cost and suitability.
getClickableElements walks the DOM tree to find interactive elements. hashDomElement uses three-layer hashing (branch path + attributes + XPath) for unique fingerprints. A 30-line iterative stack replaces recursive DOM traversal.
getMarkdownContent uses turndown to convert DOM to Markdown. getReadabilityContent uses Mozilla Readability to extract article body. Both are under 40 lines in service.ts but the design tradeoffs matter for scraper developers.
Understand browser fingerprinting and anti-detection, and see why C++-level patching is fundamentally different.
Two agents running simultaneously. One is filling a checkout form. The other takes over its tab. This isn't hypothetical — agent-browser Issue #326 describes exactly this scenario.
An IP looks reasonable. A browser profile looks normal. Code runs without errors. But put the three together and they don't align. Detection systems find inconsistency.
navigator.webdriver was defeated in 2018. 2026 anti-detection looks at mouse movement entropy, keystroke dwell time variance, sub-pixel Canvas rendering offsets.