Symbolic Coding Agents For Young Ladies

A coding agent is a chat like program for programmers, that uses multiple llm-agents in a clever way.

An llm-agent is a language model, that has been instructed to look at things in a unique way.

They are usually just called models, but a raw model needs a hat, and for clarity I will call them llm-agents.

An expensive llm-agent can be used for planning, and cheaper llm-agents can be used for coding that plan.

That is the, Orchestrator pattern.

Or smaller llm-agents can reach out for advice, to a smarter and more expensive llm-agents.

That is the, Advisor pattern.

Chaining multiple llm-agents for cleanup and verification, can make the results cleaner and more valid.

But all of this is wrong, when you want a professional application, Because a coding agent should not actually write code.

But create a symbolic or semantic instruction tree, with an ever growing library of code generating nodes.

The code generators are made by a coding agents, but they are small, well defined, and do one thing well.

Here we can use conventional AI, like goal solvers and expert systems.

And what I like to call a bookshelf of cookbooks, with recipes for the older AIs.

Yes, cookbooks and recipes must also be written, By powerful state of the art AI, they are complicated.

But you only create them once, during the development of your semantic code generator.

And you expand the library when you for example, add recipes for, a music library manger and player.

You can now reuse them in an image manger, And slideshow player.

Symbolic programming research and fruits, Would be a powerful start for your company.

Because code generation is infinitely faster, Than having a large language model creep around your code.

This is very much about building your own AI, you will still need a powerful AI, but less and less.

The way you visualize this application, is not by imagining a tree, as that is the job of the AI.

But rather, a library of application templates, similar to how word processors have templates.

AI pre-generates starting points for you, so that you never start at zero, which would be harder.

You would be given a dockview with blank panes, think outlook or photoshop, but all panes are blank.

And you have to tell a language model, to build a table of songs, and playlists.

And then instruct it, to create a playlist manager for the playlists.

A state of the art AI, will create a nicer template.

Critics of this idea, forget that they are never sure, if a feature they requested to be removed, is ever fully removed from their app.

In a stricter symbolic world, when you say remove this button, or use a UI control delete pointer even.

All the things behind it, collapse, because something that does not exist, cannot call dependencies into existence.

I think pixel-art games still require conventional agent use, for one conventional AI can’t generate sprites.

But, if you scale down the semantic programmer, to instead of creating applications, create application blueprints.

Then you may till be able to beat state of the art AI, in planning, you would create precise XML blueprints.

Based on well thought out cookbooks, but again built recipe, by recipe by state of the art AI.

This more more mechanical, rather than neural agent, holds the power to improve local AI, the smaller and more limited models.

The question here, isn’t whetter you can make a good local better, but how low can you go to have a super-fast and really small and smart AI.

And ultimately, you are just building another llm-agent, except symbolic AI is not noisy, or random, it is always extremely precise.

Go ahead, build your first symbolic AI, with a small large language model front-end.

And have your state of the art AI, consult it for precise application blueprints.