Multi Agent Workflows: Take 1
Moving onto software management with agents.
Coordinating agents is one area I'm trying to figure out how to improve. LNG CMS is designed to host websites, but it doesn't have some of the build tools you'd expect from local development, like JavaScript bundling and CSS shaking. I've been experimenting with creating a smaller Liquid template renderer that one can work through the design, and then push and pull changes between client and server.
These are different codebases, with different goals. I've gone through many iterations of my flow. It started with a bare folder, and some preprocessing, and rendering whatever liquid template matched the path in Sinatra. It's now evolved into a local YAML database, individual themes, and a full ruby API client. The client was developed by asking the local editor LLM to think about what primitives it needed, document them, and then I would manually give those directions the LNG project's LLM. This back and forth iteration, with me in the middle, helped me figure out what was possible and what was missing. I now can take a Wordpress theme, and Wordpress site export, and get a reasonable clone of a Wordpress site on LNG, albeit with some coaching of what I'd like along the way.
I've now taken the themes, and the client api, and now I've asked the local editor LLM which features are needed for MCP support. The back and forth, and LLM comments on how great the design's are, is allowing me to enhance my existing features with more powerful LLM capabilities. But it still required me to work back and forth and refine the output. And there will still be more polishing between as the features are fleshed out and used in anger.
These are different codebases, with different goals. I've gone through many iterations of my flow. It started with a bare folder, and some preprocessing, and rendering whatever liquid template matched the path in Sinatra. It's now evolved into a local YAML database, individual themes, and a full ruby API client. The client was developed by asking the local editor LLM to think about what primitives it needed, document them, and then I would manually give those directions the LNG project's LLM. This back and forth iteration, with me in the middle, helped me figure out what was possible and what was missing. I now can take a Wordpress theme, and Wordpress site export, and get a reasonable clone of a Wordpress site on LNG, albeit with some coaching of what I'd like along the way.
I've now taken the themes, and the client api, and now I've asked the local editor LLM which features are needed for MCP support. The back and forth, and LLM comments on how great the design's are, is allowing me to enhance my existing features with more powerful LLM capabilities. But it still required me to work back and forth and refine the output. And there will still be more polishing between as the features are fleshed out and used in anger.