Browser Backends¶

hafermilch supports two browser backends. Select one with the --browser flag.

Playwright (default)¶

Playwright runs a real Chromium browser as a Python-native library. It captures the page's accessibility tree via aria_snapshot() and interacts using CSS selectors and ARIA roles.

uv run hafermilch run examples/plans/saas_onboarding.yaml --browser playwright

How it works¶

1. Playwright launches a Chromium instance (headless by default)
2. aria_snapshot() on the body element produces a YAML-like accessibility tree
3. The LLM reads the tree and returns a BrowserAction with a CSS selector
4. Playwright executes the action (click, fill, navigate, …)

Strengths¶

• No extra dependencies beyond playwright install chromium
• Precise, deterministic selector targeting
• Full vision support — screenshots are sent to vision-capable models
• --no-headless flag lets you watch the browser live

Best for¶

• Products you own and whose DOM structure you know
• Authenticated flows (login, checkout) where selectors are stable
• Regression testing — comparing scores across deploys
• Environments where you cannot install Node.js

agent-browser¶

agent-browser (by Vercel Labs) is a Rust CLI purpose-built for LLM-driven browsing. Its snapshot command returns an accessibility tree where every interactive element carries a short @ref handle (e.g. @e1, @e7). The LLM picks refs directly — no CSS knowledge required.

npm install -g agent-browser

uv run hafermilch run examples/plans/saas_onboarding.yaml --browser agent-browser

How it works¶

1. hafermilch calls the agent-browser CLI via subprocess for each action
2. The snapshot command returns a tree like:

   - heading "Sign up" [ref=@e1]
   - textbox "Email" [ref=@e3]
   - button "Create account" [ref=@e7]
   

3. The LLM reads the tree and returns a BrowserAction with a @ref selector
4. hafermilch calls agent-browser click @e7

Strengths¶

• LLM never needs to guess CSS selectors
• More resilient on JS-heavy apps where selectors are generated or unstable
• Semantic, human-like navigation ("the button labelled 'Create account'")
• Session-isolated — each run starts fresh

Best for¶

• Exploratory evaluations of unfamiliar third-party products
• Products using heavy React/Vue/Angular frameworks with unstable selectors
• When the LLM keeps choosing wrong selectors with Playwright

Comparison¶

Which should I pick?¶