AI Agents and a Brand New World

I really think the future of tech will predominantly be dominated by a handful of smart people that posses carefully implemented and arranged hierarchical AI Agents at their disposal.

These people won't necessarily be AI researchers or software engineers or scientists, but instead a group of people who trained themselves over the years to master the craft: AI Agent management through very carefully crafted prompts.

These prompts will look like ordinary , weirdly organized text junks , image junks or audio junks to the average person but deep down they will be THE REAL DEAL.

I assume these instruction sets will worth millions of dollars and will be very lengthy but split into categories and priority segments.

The reason I say these are not just some idea I came up with but rather a result of my recent experiments with agentic tools and my own language model modifying and training experiments.

The amount of quality information you can generate with an LLM (or anything with a neural net really) when you build carefully designed prompts (calling it prompt libraries would be a more realistic name though) is scary.

Then there is the "connecting them together for a greater goal" part of things which is the most complex but also the one that feels like magic.

If you were to see my current software development setup on my computer, you would think I am in a sci-fi experiment or am losing my mind (both are fine by me).

Past months I have been experimenting with different ideas on making the most out of LLMs.

I  spent hours every day on crafting system prompts to make my agents smarter, for both already existing tools I use in development, such as IDEs and for my own tools such as my recent experiments on building private chatbots trained on custom data. Either way, I realized the main concept  I am realing with is literally the same:

"you either know how to guide the bot or you don't know how to guide the bot. You can either teach the bot or you can not".

You can't just write an article like this and use it as a system prompt. 

The idea behind the perfect set of system prompts is that it should be so clever and minimal that the language model will start to become and act as your perfect agent, meanwhile listening to your new instructions on inference and understanding the correlation between.

Because I am a programmer, I always have to convert my logical design of a program , into real source code so that it runs on the computer. 

This can take months.

I will now give a lot of examples from my work, but read them as if they are details of your work and you will see a pattern. I will now talk about codes, environments, linux and windows but read them as “some ground to build something that depends on another thing…”

Here we go:

Say, you are building an app and you go and tweak your diagrams, notes you have taken on a notebook, screenshots etc. then you have to go back to source code files write more code, delete more code and so on.

Pretty time consuming!

This process of converting logic into pure code in a computer prog. language like Python or C or Java is CRAZY time consuming and has tons of challenges not realted to the program but related to set of tools you are working with.

Just like trying to cook pasta but you have to find the perfect sauce ingredients from not a supermarket but a bookstore.

Opss!

One programmer example would be implementing a ping-pong simulation program you wrote (pseudo code/diagrams/equations) for both RHE Linux and Windows and one runs on x86 and other computer has an ARM chip. Now you are in BIG TROUBLE.

The program didn’t change at its core, but you have to implement the same program in VERY different ways for two different operating systems.

Will this IMPROVE THE PROGRAM? Nope.

Will it be a better ping-pong game? No.

But you have to do it anyways.

You will spend most of your time with NOT the actual logic and design of the program, but the DEPENDENCIES of it. 

These can be the RHE Linux’s ways of operating, how they handle security, graphics, terminal commands, different versioning etc.

You are doomed.

You will deal with a ton of different non-related-to-the-program aspects just so you can get the outcomes you already know.

The worst part? They will keep changing forever. The next year you compile it, everything will break.

You will spend months to even find a compatible middle ground.

This can be solved forever.

Literally.

 Building clever AI agents that can distribute themselves in an ecosystem , write platform-specific code and report back to their master agents for review and unit testing. 

Imagine an ARMY of coders. You are the programmer, the mastermind, and the coders are thousands in numbers and are super fast… imagine that!

We WILL get there, because we HAVE TO.

Most of modern coding isn't implementing your program to the actual code but rather begging the ecosystems you are in to allow you to work.

Because of what I do, I focus my efforts in codebase management and coding but the same principles apply to ANYTHING that can be done by a human and that can be communicated with words, sound or image. 

But don’t confuse automation with creation!

To automate, you must first ‘create’!

Knowing parts of the inner workings , the philosophies of theories, rules, concepts in the field are a MUST if you are trying to come up with a useful solution for a given problem. it all lies in the fundementals of the problems really. Trying to figure out a new approach? Go understand the current system and why it exists first!

So before you get excited about building an army of agents to automate things, you first must understand the current situation.

The future is crazy interesting.

We already did a lot of automation for a lot of things in the world with non-AI or semi-AI solutions, and it is about time we automate more with intelligent machines.

The scariest part? We have billions of people with near-zero innovation and creative abilities. We made people “repeat” tasks and told them this is the way to be productive. The world needed people to repeat tasks and awarded them for it.

Not the case anymore.

Times are changing RAPIDLY.

Very soon, we will realize the importance of “creating, innovating, and automating”.

The next time you read a technical article, for god’s sake please don’t try to memorize it and repeat it to yourself like a parrot would.

Do this thought experiment instead: “I read this, it has a point, did I understand the point of this concept? Ok, I think I did. Now, I should think of a good question to ask myself about this. Why? Because I want to grasp the idea better. Why? So that I can come up with something new and useful in the near future. Hm. Let me read it once more to make sure.”

You better get used to it. This is how you make yourself productive. Don’t believe me? Look at the thought process of an LLM that excels at reasoning. They are meant to mimic human creativity! So, if you haven’t been doing this already for long years, you missed A LOT.