Why enterprises should stop attempting to automate out engineers in one big shot and allow engineers to naturally build their own automations.

[Merzal#1414]

Why enterprises should stop attempting to automate out engineers in one big shot and allow engineers to naturally build their own automation tools to become more and more efficient over time until the engineer becomes an artificial intelligence operator before the operator then also evolves into something else.

Enterprises may be looking at agents as a big juicy new technology which on paper gives enterprises the opportunity to replace roles with agents, let me explain why this for now, is not going to happen the way you think it is.

Firstly, an agent is just a fancy term for at a basic level, an LLM with some pre context, in the medium stage, an LLM with pre context and a vector database while reading from files, the internet or traditional databases and at an advanced level, can perform certain tasks like triggering pipelines and clicking buttons on dashboards based on outputs.

This is not a new technology, it’s a name which start ups are using in order to package LLM’s into something they can sell to enterprises, the name stuck around and now it’s a “thing”.

The concept is exciting for people, an agent is a better way to provide a vision where something is actively performing jobs.

The reality right now is very different to that vision, the reality is that LLM’s still often provide strange responses, vector databases are not a sure-fire way to “teach” an LLM about your business or your codebase, for example, has anyone noticed that LLM’s are very bad at NOT doing the things you don’t want them to do?

That’s because when you tell an LLM to not do -insert something here- then you are technically teaching it to remember the very thing you don’t want it to do which eventually leads to overfitting. This is just an example of a limitation where people have a big misunderstanding of how LLM’s work which originally are just predictive systems which use zeroes and ones to determine which word should be in a sentence.

“I need to feed my _____ in my aquarium” the LLM determines the word is fish based on similar sentences it has been trained on by doing what’s called transforming which is where words become zeroes and ones which makes it easier to make comparisons of the importance of words by first tokenising the words and then embedding them into vector space before comparing positions of other embeds, there is no magical intelligence happening here, yet.

Long story short, an LLM is just predicting the next word based on the words you’ve given it in the past, that means if you want an LLM to not do certain tasks, you have to provide the tasks it should not do and therefore you are working against yourself because it now knows about those tasks and has a greater chance of mentioning them, if you don’t understand the problem here, I suggest you further look into it because it can open your mind in terms of understanding that an LLM is over estimated in this area. The reason why it feels so “intelligent” is because of the scale of trainings, we are talking about trillions of parameters.

GPT-2 (2019) – 1.5 billion parameters
GPT-3 (2020) – 175 billion parameters
GPT-4 (2023) – Estimated in the trillions (not officially disclosed)
Google Gemini 1.5 (2024) – Likely over 1 trillion parameters
Human Brain - 86 billion neuron, with each neurone connecting to other neurone resolution in over a quadrillion connections.


And look at how quickly it’s growing! Yet there is something that these companies know and fear. A fact, a mathematical equation which means that these LLM’s are not just going to infinitely scale forever and get “smarter” with absolutely giant investments in space and resources.

There are laws which govern this like Chinchilla’s Law

L = Loss (error of the model)
N = Number of parameters
D = Number of training tokens
a, b = Constants (empirical values: a ≈ 0.34, b ≈ 0.28)

For models to be balanced as they grow, the infrastructure and electricity requirements become exponentially increased, sure, Deepseek managed to make an LLM based on previous LLM history, but we all know what garbage in garbage out means here, it means that Deepseek isn’t an upgrade, it’s more of a copy, you can adapt an LLM to make it better with that strategy but you aren’t necessarily moving the LLM technology forward in terms of what’s possible.

This brings me to vector databases and all of the different companies offering you this -> “give all of your business data to us and we will train an LLM to know everything about your business!” Where the offer is awesome, the reality isn’t as good, first of all, this isn’t training and once you add ALL of your company data into a vector database, you are going to start increasing the chances of the LLM responding with company data which is unrelated to your question, for example, if you have added a bunch of data about your business to an LLM and then asked the LLM about where you should walk your dog next, it might just tell you to walk your dog on confluence! It won’t make any sense.

The purpose of me writing this post is to protect people from having massive expectations and ending up wasting a lot of time and money to be disappointed, I believe businesses should allow their employees to grow with AI and it will be very clear where time to result is being reduced, programmers are obviously very good at automating their own tasks, they are the ones making the AI after all, they are the ones making all of the tools and they are the ones benefitting the most out of becoming much more efficient through the use of AI, a lot of professions and industries still see barely any use for AI, do you see a brick layer asking chatgpt where to put the next brick? -

No, a brick layer is going to be a lot less adept at using any form of AI compared to a programmer who uses AI on a daily bases consistently and prominently throughout the day. - Think someone who isn't a programmer will be able to utilise and manage an Agent better than a programmer? Think again, think agents won't need management? Think again, we need to manage humans, we need to not only manage agents but we also need to design, architect, build, maintain and update agents.

Programmers know AI best and they will be the ones who will automate themselves over time and the reason is simple, human brains are efficiency machines, they will always look for the most efficient and easy path forward, for programmers, that path currently is to automate things they do through AI, thus, all you need to do, is enable your programmers to naturally build tools to make their own life easier, to make your business more efficient and over time you may see that the role of a programmer completely changes as they need to do the manual labour of typing out code less and less and less based on their own developments.

This is the same as a fisherman who buys a boat and casts a giant mechanically powered net instead of trying to catch fish with his hands, who do you think is figuring out how to catch more fish? The fishermen themselves of course! Do you need less fishermen this way to catch the same amount of fish? Yes, but you could also just catch more fish and my final point here is, you need a fishermen to run the boat because he understands the fish the most!

Architects are well positioned to benefit a lot from AI too, in fact I thought it was a dream tool and yet I am now convinced that architecture is a lot easier to automate than programming too, because in order for services to be created automatically, the architecture needs to be understood and generated first, there is also a lot less input and output required to form flows compared to creating working business logic.

In terms of agents, it’s clear that start ups took advantage of repackaging LLM’s to sell them to enterprises but now larger players will create the best agents making it difficult to compete with, unfortunately knocking out the majority of agent based businesses, will agents quickly replace programmers, designers, marketing?

Let’s be real, the agents will help the programmers and designers and marketers, as they help more and more, the human will have to do less and less and their job will evolve into something new, essentially a manager of agents. However to think enterprises will simply deploy an agent some time this year and that agent will replace a capable software engineer simply isn’t going to happen, it’s the software engineers themselves that will need to make this happen as they themselves know best what such an agent needs to do.

This opens up an interesting dilemma, managers who do not understand Artificial Intelligence, or how to create and operate good agents, will be much less resourceful because the future of enterprise is headed towards a smaller number of people managing both people and agents.

We can see that from every agent based platform on offer, there are several agents working together to accomplish goals, this means that agents are purpose built and will need to be constantly developed, usually an agent is just pre context and a place to get some more pre context from (vector database, internet, traditional database, documents), but to get real value from an agent, the agent needs to perform operations like clicking buttons and triggering something within a business.

We can’t pretend that these agents will not be difficult to build, the dashboard to operate them needs to be considered and the overall design of the agent flow has to be created. Who will do these bits and pieces? Another Agent? Maybe if the agent exists in a Quantum computer.