I’ve been working on DaveLovable, an open-source experiment around AI agents for web development.
The core idea: Instead of a single LLM generating code, I use a small multi-agent system (Planner + Coder) that can:
* generate React apps * modify existing code * run terminal commands * commit changes to git automatically
Everything runs with a real Node.js environment in the browser using WebContainers, so the preview is not mocked.
I also added a visual editing mode (click on UI → modify styles) and multimodal input (you can upload mockups or PDFs and ask the agent to recreate them).
One interesting challenge was coordinating agents reliably without them getting stuck or over-planning.
Still early, but I’d love feedback, especially from people working on dev tools or AI agents.