An idea: Rough draft

An idea came to me recently: how do we get Agents(or what I affectionately call “Clankers”) to write more reliable software and make more reliable decisions in a more explainable way? Before I dive into the idea, I want to give some context to build up the idea so that for those of you who are reading who have no background can understand what I’m doing. How I arrive at the idea is sort of also fundamental to the idea as well.

The throughline of this draft is roughly, we want agents to do tasks or come to some outcome. In order to do that, we need to:

Each step in the chain, I have an intersting core idea that will be explored in future blog posts. For now I just wanted to write down a simple core idea to share with colleagues and friends.

Worked example: Baking Pie (or making the best pie)

This is the perfect example for demonstrating this idea!

Imagine you’re at a strawberry pie shop making the best strawberry pie. How might you do it if you were in charge of this? You might first start with a recipe for making pie, which would be represented by this diagram:

Figure 1: A recipe for strawberry pie, drawn as a string diagram: ingredients are wires that flow through processes (boxes) into the finished pie

Your pie needs to exist within certain constraints. For example, because you picked strawberry pie, you know that you aren’t going to be sourcing oranges when looking for ingredients(I’m trying to invoke a “type system” here for those that get it). However, within the space of valid ingredients, we might have some cost function to try to pick the “best” strawberry + flour + some other ingredient to make the best pie! In jargon, you might use a type system to eliminate invalid combinations of ingredients, and some sort of cost function or SMT solving to figure out of the valid combinations what is the “best” combination for your shop.

To come up with the best filling strategy - we might have two “agents” working together, one taster agent that checks, given some hidden criteria, that the recipe is “good” and one maker agent that comes up with recipes based on the information they recieve from the taster. These two agents would have to communicate to each other on what the best pie recipe is. They might have these two threads(back and forth conversation):

MakerTaster
I’m going to make strawberry filling. I’m going to try to make it really sweet because I like sweet and Americans like sweet. I’ll put in 90g of sugar and 10g of strawberries.
→ sends recipe to taster
I think Americans tend to make things really sweet, and you can’t taste the nuance in the strawberries. This strawberry filling is too sweet!
← sends verdict to maker
This taster sucks. He doesn’t like my sweet filling. Is he some sort of European? “What do you mean by that?”
→
Well, I generally think that although Americans like sweet things, strawberry fillings should be less sweet to bring out the nuance of the strawberry.
←
Ok. Let me reduce the sugar to 40g and increase strawberries to 60g. sends new filling to taster
→
Much better!

Result: Good Filling.

This sort of message passing is codependent and can’t be represented by a string diagram, but does encapsulate a specific process in which a process for making something is refined.

Obviously, there are only so many agents in your fictional pie shop! This requires a scheduler to pick which order in which the agents perform tasks, and agents often times need to context switch between tasks as well! Obviously in the agentic world we can design harnesses to make the agent specialize in a specific field and then run the agent in machines distributedly!

This is a dummy example, but we can imagine that this process is going on right now in any decently large organization. Thus, why can’t we apply this idea into our Agentic AIs as well to get the best result out of these clankers and be a mini CEO of our vibe coded apps?

I will go into depth with some of these ideas in smaller blog posts in a bit (not to any formality), but just enough to outline the idea. Many of these ideas were inspired by reading about string diagrams and “the society of mind” and I will reference them here in the references.

Some of my References

Just so you know, I just read the table of contents of most of these and didn’t really read them through ;)

Working on this

I will be working on this sporadically as this is my side quest and not my main goal. My main goal is.. to get a girlfriend. It’s a rough and long ever lasting journey without end.