Building the Missing Standard for Agentic Workflow Orchestration

Hacker News

Share card

Building the Missing Standard for Agentic Workflow Orchestration

I built Grapheteria - https://github.com/beubax/Grapheteria . Yes, I know what you're thinking "Oh no, not ANOTHER agentic workflow library" I felt the same way, but hear me out on why I think we still haven't hit the sweet spot. We've all been caught between two frustrating options - Code-only frameworks: Powerful but often buried under layers of abstractions UI-only builders: Great for simple flows but hit a wall when you need real customization Here's my take: Code is non-negotiable. No UI, no matter how good, can replace the flexibility, version control, and deployment options that code provides. But a good UI is invaluable for visualizing your flow while you build with code. Seeing what you're creating helps catch logic errors early and makes complex flows manageable. Building agentic workflows doesn't need to be complicated. A few key features should work out of the box: - Human-in-the-loop - Step-by-step debugging - Solid logging Everything else should be up to the developer to add as needed. Grapheteria follows a simple principle: design clean, composable graphs where each node and edge has clear purpose. Edit in code, see it in UI instantly. Edit in UI, code updates automatically. Grapheteria never restricts your ability to customize. Every aspect remains accessible through code, while the UI provides immediate visual feedback. Key characteristics: - Zero abstraction tax - the code you write is the code that runs - Pass any data type between nodes (even ML models or Redis queues if you want to), not just strings - Visually debug your flows via the UI These become building blocks for larger systems. Agents can dynamically modify the workflow at runtime whether it be to add new agents or change paths. Deploy your Grapheteria flows in accordance with the A2A protocol and embrace a truly asynchronous multi-agent orchestration. I believe agent systems work best when built from small, specialized state machines rather than monolithic agents trying to do everything. When your workflow is a well-defined graph, both reasoning and execution become transparent. Check it out here: https://github.com/beubax/Grapheteria Thank you for your time! What are your thoughts?