Hi HN,
Agents can nowadays design UI very competently, but the wider question of graphic design is largely unsolved.
On one side, you have image models like ChatGPT images, which are high-quality but recognizable and barely editable. On the other hand, you can use LLMs and accept results that, well, look like UI.
I've been working on Rivermark for a few months now to tackle this issue. It's an infinite canvas with an agent that excels at creating on-brand visual content for socials, marketing, and to some extent print. The agent is extremely context-aware and fast (we call it a "short horizon" agent). It will ingest anything and run with it - from Google Sheets copy/paste to .zip files with brandkits or photo sets. Everything is editable manually like in Figma too.
We found it works best for people who are not designers but who know what looks good. MCP features to connect it to your agents are coming soon.
We're powered by superfast inference at almost 1000 tokens per second and a custom realtime harness, meaning the agent is non-blocking and always responsive, and even has a non-turn-based voice mode that you can just talk into.
We have some teams and early power users shipping work with it and are learning a lot. There have been numerous engineering and product challenges to even get here - I'd be happy to talk about them!
Comments URL: [https://news.ycombinator.com/item?id=50364525](https://news.ycombinator.com/item?id=50364525)
Points: 1