LangChain Optimizes Browser Use for Deep Agents with Stagehand, Cutting Costs by 13x

Stagehand toolset achieves 90% success rate on WebVoyager at a total cost of $1.98, compared to Playwright MCP's 78% success rate and $26.71 cost (13x difference).
Code mode allows agents to write code to execute multiple browser steps in a single tool call, reducing average steps per task from 26 (Playwright MCP) to 10 and median navigation response size from 17,000 characters to 117 characters.
Snapshots compress page accessibility trees into text with IDs, enabling direct actions like {"op": "click", "id": "0-7"} without raw HTML or selectors, improving token efficiency by 80% compared to Playwright.
Settle wait optimization replaces fixed 8-second sleeps with a 500ms quiet period check, reducing median agent time on HardBench from 182 seconds to 124 seconds without accuracy loss.

Toolset Comparison

Benchmarks include WebVoyager (100 tasks, Claude Sonnet 5.5), HardBench (102 tasks, Claude Opus 5.5), and WebTailBench (150 tasks).
On HardBench's 38 core tasks, Stagehand's run tool passed 30 tasks, while Browserbase's browse CLI passed 21 tasks with a 3x cost difference.
Restricting Stagehand to its native API only (disallowing Playwright-style code translation) decreased success rate by 5.3 percentage points and increased cost per passed task by 19%.

Stagehand v4 runs as a browser extension to handle low-level calls near the page, making it 2x faster than Playwright in Browserbase benchmarks.
The settle wait fix was upstreamed to Stagehand via Pull #3282.
Browser use is available in Managed Deep Agents and can be added to open-source Deep Agents using StagehandMiddleware from langchain-stagehand.