Inspiration
Building SEO backlinks is traditionally a manual, tedious, and soul-crushing process. We wanted to move beyond AI tools that just give text advice and build a true agent that acts as your hands on the web. The goal was to create an autonomous system that can visually navigate, research, and execute outreach just like a human marketer would.
What it does
AgentLink is an autonomous SEO campaign manager that orchestrates a team of specialized AI agents:
The Profiler: Analyzes your website to generate a dynamic, comprehensive Master Company Profile.
The Researcher: Scours the web to find highly relevant niche directories and blogs, piping them directly into a real-time Kanban board.
The UI Navigator (Form Filler): Via a custom Chrome Extension, this agent reads the DOM state of target websites and takes over your browser to physically type, click, and navigate complex submission forms.
The Pitcher: For blogs and bespoke sites, it autonomously drafts highly personalized email pitches ready for outreach, using relevant information from both our website and the target site
The Extractor: If the Form Filler encounters a field missing from your profile (like a local zip code), it pauses, calls the Extractor to hunt down the missing data on the target site, and seamlessly resumes filling the form.
How we built it
We built a modern, serverless architecture centered around the Google Agent Development Kit (ADK):
The Brain: Google Gemini 3 powers the reasoning, orchestrating ReAct loops via our Go backend.
The Hands: A custom Chrome Extension parses interactive DOM elements into a simplified JSON state and executes standard JavaScript commands dictated by the Gemini model.
The Interface: A React (Vite) SPA utilizing Tailwind CSS and dnd-kit provides a premium Kanban board, synced in real-time via a local Firebase/Firestore emulator.
Challenges we ran into
Bridging the gap between an LLM's reasoning and physical browser execution was difficult. We had to design a robust Chrome extension capable of extracting interactive elements without getting bogged down by massive HTML payloads. Mapping temporary IDs to DOM elements so the agent could accurately target and manipulate them took significant iteration.
Accomplishments that we're proud of
Achieving true Agent-to-Agent communication. Seeing the Form Filler agent realize it doesn't know a specific detail, autonomously delegate a search to the Extraction agent, and then use that new data to successfully submit the form is a massive step beyond standard text generation. We are also incredibly proud of how smoothly the Go ADK handles these complex ReAct loops.
What we learned
We learned that LLMs are incredibly capable of UI navigation and complex workflow execution if you provide them with a clean, structured state. We also discovered that breaking down a massive task (SEO outreach) into specialized, delegating sub-agents (Research, Fill, Pitch, Extract) yields much higher reliability than trying to use one massive prompt.
What's next for AgentLink
We plan to expand the multimodal capabilities to use raw screen captures (vision) instead of just DOM parsing to bypass shadow DOMs and complex anti-bot measures. We also want to integrate SMTP APIs to fully automate the dispatch process for our Pitch agent's drafted emails.
Log in or sign up for Devpost to join the conversation.