Stagehand

Quick Take

Stagehand is Browserbase’s open-source SDK for browser agents: a framework that lets you write browser automation using natural-language actions, deterministic code, or a mix of both in the same script. Version 4, released in August 2026, adds a browser-extension execution mode, self-healing actions, and result caching, with Browserbase reporting it runs roughly twice as fast as Playwright while using around 80% fewer tokens.

What Stagehand Does

Stagehand exposes a small set of core actions for controlling a browser with AI assistance:

• act(): perform an action described in natural language, like “click the submit button” or “fill in the search box with ‘AI tools’”
• observe(): ask the model what actions are available on the current page before deciding what to do
• extract(): pull structured data out of a page using a described shape rather than manual selector-writing

The key idea is that you don’t have to choose between full AI autonomy and traditional, brittle selector-based scripts. You can write exact Playwright-style code for the stable, well-understood parts of a flow, and drop into natural-language act() calls for the parts that are more dynamic or where writing a precise selector would be fragile.

What’s New in v4

• Browser-extension mode: Stagehand now runs its logic inside the browser as an extension, bringing the “local” execution experience to remote browsers too, which lowers round-trip time and reduces potential race conditions with the Chrome DevTools Protocol
• Self-healing actions: automations adapt when a page’s layout changes, instead of breaking outright
• Improved iframe support and better context management overall
• Caching for act(), observe(), and extract(): once identical results are seen enough times (a configurable hit-count threshold), Stagehand serves the cached result and skips a fresh model call entirely, cutting both latency and cost on repeated runs

Key Facts

• Developer: Browserbase
• License: MIT, fully open-source
• v4 released: August 2026
• Model support: OpenAI, Anthropic, and Google Gemini, via the Vercel AI SDK; Browserbase’s Model Gateway lets you access all supported models through a single Browserbase API key at the same price as going direct to the provider
• Performance (Browserbase’s benchmarks): roughly 2x faster than Playwright, about 80% more token-efficient
• Ecosystem: partnered with Vercel, Mastra, CrewAI, and LangChain on browser-agent templates and integration guides

Where Stagehand Fits in Ainanza’s Browser-Agent Cluster

Stagehand sits at the developer-framework layer of browser automation, alongside Browser Use, which offers a similar open-source, model-agnostic approach but without Stagehand’s specific code-plus-natural-language blending. It’s a different layer than Cloudflare Kitesurf, which is a lightweight browser runtime rather than agent logic, and different again from full computer-use agents, which control an entire desktop rather than scripting inside a browser session. Traditional Playwright automation, which Stagehand is commonly benchmarked against, requires hand-written selectors for every interaction and breaks when a page’s layout changes; Stagehand’s natural-language actions and self-healing behavior are built specifically to reduce that fragility.

Honest Limitations

• Requires programming knowledge. Like Browser Use, this is a developer SDK, not a no-code tool.
• Model API costs still apply. The SDK is free, but every act(), observe(), or extract() call (when not served from cache) uses a paid model call.
• Reliability depends partly on the underlying model. Natural-language actions are only as good as the model interpreting the page and deciding what to do.
• Newest features are, well, new. v4’s extension mode and caching are recent additions; expect some rough edges as they mature in production use.

Alternatives Worth Knowing

• Browser Use, an open-source, model-agnostic framework with a similar goal, without Stagehand’s code-plus-natural-language blend
• Cloudflare Kitesurf, a lightweight browser runtime rather than agent logic, a different layer of the same stack
• BrowserOS neo, a local, open-source browser built specifically for AI agents to drive
• BrowserAct, browser automation aimed at less technical users
• Cua, computer-use agent infrastructure that controls a full desktop, not just a browser

For the concept behind all of these, see the browser agent glossary entry, and see Cloudflare Kitesurf vs Browser Use for how the runtime-versus-framework distinction plays out in practice.