gmanu0077/autonomous_web_agent_

Web Agent Framework, or WAF, is a three phase autonomous agent that controls a real Chrome browser. Given a URL, it can either execute actions on the page, such as clicking buttons, filling forms, and navigating, or answer questions about the page's content. The distinguishing approach is that it does not simply dump raw HTML into an AI model. Instead, it builds a structured graph of every element on the page and queries that graph to find precise selectors before generating any automation code. Phase 1, page preparation, handles the grunt work: dismissing cookie banners and popups, scrolling to reveal hidden content, and building a graph of every DOM node on the page, which is then indexed into a vector store. Phase 2 detects whether the user's request is an action or a question. Phase 3a, action mode, uses the DOM graph to find relevant selectors, asks an LLM to generate targeted Python code using Pydoll or Selenium, runs it against the live browser, then verifies the result and retries if it failed. Phase 3b, chat mode, answers plain language questions about the page using semantic search over the indexed DOM nodes. The AI backbone supports Google Gemini or a locally running Ollama model. The pipeline is built on LangGraph. The project is written in Python. The full README is longer than what was shown.