Hi HN! I've been working on a new type of LLM agent framework designed with consistency and observability in mind. The agent behavior is represented as a directed graph, you can save/load agents, and the agent representation is JSON so you can store agents or send over the wire. It's also designed to have modular behavior that can be customized, and makes it easy to connect to real-world systems.

Would love to hear anyone's thoughts!

Repo: https://github.com/anerli/langur There is a Discord as well: https://discord.gg/wSBSP56V7U