LangChain details how it cut browser-use context bloat for deep agents with Stagehand
hwchase17 · x · 2026-11-04
LangChain's Sydney Runkle (RT'd by Harrison Chase) published "How We Optimized Browser Use for Deep Agents," examining why browser use is brutal on agent context windows: tasks are long-running with tons of sequential tool calls, and web pages carry lots of task-irrelevant content.
Key points:
- The team chose Stagehand from Browserbase, a toolset purpose-built for agents navigating the web, over generic browser automation.
- The article walks through the full selection process and evaluation criteria.
- Crucially, they mined suboptimal agent trajectories in LangSmith traces to find where context was wasted, then upstreamed those optimizations into the Stagehand toolset itself.
A rare engineering retrospective where trace data drives tool-level design improvements.
Related event: LangChain Explains Browser Tool Selection for Deep Agents(2 posts)→
More from coding & agent
- Sentry's AI Agent goes head-to-head with Semgrep bot in a real PR review — zeeg · 2026-11-04
- Sentry Engineer Shares Bots Arguing on a PR: Security Bot Flagged, AI Bot Refused and Resolved — zeeg · 2026-11-04
- Zed's zeeg: Sentry tools now globally disclosed and already shipped — zeeg · 2026-11-04
- Particle-1 lands in OpenClaw with ~45% fewer output tokens than DeepSeek V4.1 Flash — garrytan · 2026-11-04
- Grok bots vs OpenClaw: multi-LLM agents race to forecast US elections live — alvelda · 2026-11-04
- Capy makes BYOK free on every plan, supporting Claude Code, Codex and 20+ providers — garrytan · 2026-11-04