Agents Need Hands, Not Just Brains — How RPA Fills the Automation Gap

Today's GitHub Trending tells a clear story: agents are moving from generating content to operating real software. Three repos made this obvious: Chrome DevTools MCP (ChromeDevTools/chrome-devtools-mcp) gives coding agents direct access to browser debugging via the DevTools protocol. Agents can now inspect DOM, monitor network requests, and debug their own output — inside the browser. Univer (dream-num/univer) packages Office capabilities (editing, calculation, collaboration, file processing) as an SDK, explicitly designed for "people and AI agents." You embed it into your product, and agents can read/write spreadsheets and documents. plannotator (backnotprop/plannotator) builds a workspace for reviewing agent output — shape, decide, plan, annotate, review, and send feedback to agents with one click. Together, they reveal a pattern: the community wants agents to operate real software, not just generate text. The Gap Nobody's Addressing But both Chrome DevTools MCP and Univer have a limitation: they require either a specific environment (browser) or SDK integration into your product. What about the software enterprises already run? Legacy ERP systems with no API Internal OA portals that only work in IE mode Desktop finance software with no SDK Web apps behind authentication that you can't modify You can't integrate an SDK into software you don't own. You can't use browser DevTools for desktop apps. The Third Path: RPA RPA (Robotic Process Automation) takes a fundamentally different approach: it controls software from the outside, simulating human interactions — clicking buttons, filling forms, reading screens, navigating menus. No SDK integration. No code changes. No API requirements. iflytek/astron-rpa (https://github.com/iflytek/astron-rpa) is an Agent-ready RPA suite that brings this capability to AI agents. It's open-source, works out of the box, and is designed for both individuals and enterprises. Here's how it compares to today's trending repos: Approach Chrome DevTools MCP Univer SDK astron-rpa What it controls Browser debugging Office documents Any desktop/web software Integration needed MCP protocol SDK embedding None (external control) Best for Frontend debugging Building AI office apps Automating existing systems The Full Loop: Brain + Hands RPA alone is just hands. You also need a brain to decide what to do, plan the steps, and monitor results. That's where iflytek/astron-agent (https://github.com/iflytek/astron-agent) comes in — an enterprise-grade agentic workflow platform for building SuperAgents. The combination: Agent plans: Break down a complex task into steps, decide which steps need RPA execution RPA executes: Operate the actual software — fill forms, scrape data, run approval flows Agent monitors: Check execution results, decide next step, handle exceptions Human reviews: At critical nodes, pause for human confirmation before continuing This mirrors what plannotator showed us today: agents need human oversight. But instead of reviewing code output, we're reviewing software operations — did the form get filled correctly? Was the right data scraped? Should we proceed to the next step? Why This Matters Now Today's trending repos prove the community is ready for agents that operate real software. But the paths being built (browser MCP, Office SDK) only cover greenfield scenarios where you control the software stack. For enterprises running legacy systems they can't rewrite, RPA is the bridge. And making RPA Agent-ready — so agents can call RPA operations as part of their workflow — is the missing piece. That's exactly what iflytek/astron-rpa delivers. Links: iflytek/astron-rpa: https://github.com/iflytek/astron-rpa iflytek/astron-agent: https://github.com/iflytek/astron-agent