LLM-Powered RPA

LLM-Powered Robotic Process Automation

# Traditional RPA: brittle, breaks on UI changes
# LLM RPA: understands intent, adapts to changes

import anthropic
from playwright.async_api import async_playwright
import base64

client = anthropic.Anthropic()

async def llm_rpa_agent(task: str, url: str):
    """LLM-guided browser automation."""
    async with async_playwright() as p:
        browser = await p.chromium.launch(headless=True)
        page    = await browser.new_page()
        await page.goto(url)
        history = []

        for step in range(20):
            # Capture current state
            screenshot = base64.b64encode(await page.screenshot()).decode()
            dom_text   = await page.evaluate("document.body.innerText")

            # Ask LLM what to do
            resp = client.messages.create(
                model="claude-opus-4-5", max_tokens=500,
                messages=[{"role":"user","content":[
                    {"type":"image","source":{"type":"base64","media_type":"image/png","data":screenshot}},
                    {"type":"text","text":f"Task: {task}
History: {history}
Dom snippet: {dom_text[:1000]}
Next action (CLICK css_selector | TYPE css_selector text | SCROLL | DONE):"}
                ]}]
            )
            action = resp.content[0].text.strip()
            history.append(action)

            if action.startswith("DONE"): break
            await execute_browser_action(page, action)

        return history