managed-browser 0.2.0

Allows seamless interplay between agentic browser-use and Playwright internals

Managed Browser

A minimal wrapper to seamlessly integrate browser-use agents into your existing Playwright (Python) workflows.

---

Motivation

Browser agents powered by LLMs can think creatively—click links by intent, extract arbitrary data, and navigate dynamic UIs—but they remain inherently non-deterministic. They may:

• Misinterpret a button label or page layout
• Click the wrong element when the UI changes
• Drift off-task if given an open-ended prompt

Meanwhile, your Playwright scripts are rock-solid and repeatable, but rigid for anything beyond well-defined interactions.

managed-browser bridges the gap:

1. Drop-in Agent Segments
   Invoke an LLM-driven agent exactly where you need it, without refactoring your existing Playwright logic.

2. Seamless Control Transfer
   Hand off to the agent, then regain the exact same Page—with cookies, localStorage, DOM state, and session intact—so you can resume deterministic flows.

3. Best of Both Worlds
   Let Playwright handle the mechanical, repeatable steps; delegate the fuzzy, creative extraction to your agent, then pick up where you left off.

4. Full Visibility & Audit
   Wrap every agent run in a Playwright trace ZIP for step-by-step replay, screenshots, network logs, and DOM snapshots.

---

Installation

Requires Python 3.11+.

pip install managed-browser

---

Quick Example

import asyncio
from browser_use import BrowserConfig
from langchain_openai import ChatOpenAI
from managed_browser import BrowserManager

async def main():
    bm = BrowserManager(browser_config=BrowserConfig(headless=True))
    llm = ChatOpenAI(model="gpt-4o")

    async with bm.managed_context(use_tracing=True, tracing_output_path=...) as session:
        page = await session.browser_context.new_page()
        await page.goto("https://example.com/product/42")

        # 1) Agentic handoff: extract product info into JSON
        agent = session.make_agent(
            start_page=page,
            llm=llm,
            task="Return JSON: {\"title\": <product title>, \"price\": <price as number>}."
        )
        product_info, page = await agent.run()
        print("Extracted:", product_info)

        # 2) Hand control back to Playwright for a follow-up action
        await page.click("button#add-to-cart")
        await page.wait_for_selector(".cart-count")
        print("Added to cart, cart count:", await page.inner_text(".cart-count"))

    await bm.shutdown()

asyncio.run(main())

---

Core API

BrowserManager(browser_config, *, autostart=True)

• .managed_context(use_tracing=False, tracing_output_path=None, randomize_user_agent=True, context_kwargs=None)
  Async context manager yielding a ManagedSession.
• .shutdown()
  Gracefully close the browser.

ManagedSession

• .browser_context → Playwright BrowserContext
• .make_agent(start_page: Page, llm: BaseChatModel, **agent_kwargs)
  Returns an AgentWithControlTransfer whose .run() yields (output, final_page).

---

License

MIT © Amrit Baveja