Let AI drive the browser, but cache the route to save tokens

What it does

Stagehand is a browser automation framework built on a CDP engine that lets you steer browsers with a mix of natural language and deterministic code. You can click elements, run multi-step agents, or extract structured data into Zod schemas, all within the same script. The goal is to keep the flexibility of AI-driven navigation without surrendering the predictability you need in production.

The interesting bit

The framework treats LLM calls as expensive exceptions, not the rule. After an initial AI-driven run, it auto-caches the resulting actions and replays them without inference; when a website changes and the cached path breaks, self-healing triggers a fresh LLM consultation to patch the workflow. It is essentially a just-in-time compiler for browser automations that falls back to an AI developer only when the DOM shifts.

Key highlights

• Hybrid control: switch between direct commands (act()), autonomous agents (agent()), and structured extraction (extract()) in one flow.
• Token economics: caches repeatable actions after the first LLM run, skipping inference on subsequent executions.
• Self-healing: detects broken automations—presumably from DOM changes—and re-engages the AI to repair the sequence.
• Schema-aware scraping: extract() returns typed data using Zod schemas, not raw text dumps.
• External dependencies: the project is “best when” paired with an LLM provider API key and Browserbase credentials.

Caveats

• It is not designed to run fully offline or vendor-agnostic; the README explicitly recommends both an LLM API key and a Browserbase account for proper use.
• The mechanics of “self-healing” and auto-caching are described at a high level; there is no detail on detection thresholds, cache invalidation, or fallback costs.

Verdict

Grab it if your scrapers constantly break and you would rather pay occasional token fees than rewrite selectors every quarter. Pass if you need a completely self-hosted, deterministic automation stack with zero external AI dependencies.