Agency

Inspiration Building an AI agent workflow today means writing glue code. You stitch together prompts, API calls, and output parsers — and if anything changes, you rewrite it all. There's no visual way to compose agents, no way to see the flow, no way to reuse a pipeline you built once. AGENCY was built to change that: a visual node-based editor where you drag, connect, and configure AI agents into automated workflows — no glue code required.

What it does AGENCY lets you compose AI agents into executable workflows using a visual canvas. Drag-and-drop skill nodes (Brand Architect, UI Artisan, Code Forger, Content Weaver, Reviewer, Hardener), connect them visually with outputs flowing automatically between agents, and execute with one click. The DAG resolver runs agents in parallel where possible, sequentially where needed. Real-time WebSocket streaming lets you watch each agent work node by node. Save workflows and reuse them with different inputs. Example pipeline: Brand Architect → UI Artisan → Code Forger + Content Weaver → Reviewer produces a complete landing page.

How we built it Backend is Go with goroutine-based concurrent agent execution and WebSocket streaming for live node status. Frontend is React 19 + TypeScript + Vite with React Flow for the visual canvas, Zustand for state management, and Tailwind CSS v4 for styling. The execution engine uses topological sort on a DAG to resolve dependencies and run nodes in parallel where possible. The Go backend also integrates with OpenRouter API (Gemini 2.0 Flash) for real agent execution, with fallback template outputs when no API key is provided.

Challenges we ran into Canvas performance at scale — React Flow struggled with 50+ connected nodes, requiring virtualization and level-of-detail rendering. DAG execution with WebSocket feedback while maintaining correct topological order required careful goroutine management. Making each skill node's configuration form intuitive without cluttering the canvas was a constant design tension.

Accomplishments we're proud of A fully functional visual node editor where you can build, run, and save real AI workflows. The Go backend handles concurrent agent execution with millisecond-level streaming. The "Build a Landing Page" workflow actually produces deployable output. Clean, professional UI that looks like a real product, not a hackathon prototype.

What we learned Visual programming for AI is the future — dragging nodes is 10x faster than writing glue code for common patterns. Go's goroutine model is ideal for concurrent agent execution. The business model matters even at a hackathon — designing monetization alongside architecture forced us to think about scalability from day one.

What's next Skill Marketplace with community-shared nodes and 70/30 revenue split. Custom node SDK for anyone to build their own skill nodes. Workflow templates for common use cases. PostgreSQL backend for saved workflows and teams.

Built With

• framer-motion
• go
• openrouter-api
• react
• react-flow
• tailwind-css
• typescript
• vite
• websocket
• zustand