AILab Howest

The first point we will look at is the definition of the word 'Agent'. For many, an Agent is a system that takes action on its own. Of course, there is always some kind of starting point or trigger. This can be receiving an email, a value on a sensor that becomes too high, or a simple start button. You can then ask yourself the following question.

What is the difference between Agents and workflows?

I like to use an analogy for this. A workflow can be seen as a kind of 'train track'. These arevastpaths that follow fixed routes. This does not mean that there cannot be AI or Language Model involved!

Take, for example, the automatic summarization of an email and placing it in a document. This is an example of a workflow where one of the stops is an LLM for the summarization.

An AI agent can be seen as a taxi with a GPS that adapts to the situation. There is not one specific route; here the language model will make a decision on which route will be taken.

Here you can think of a solution that, instead of summarizing an email, will analyze whether it is a positive or negative email and then decide to take action (we only provide the options or locations, but not the paths or tracks).

Hopefully, this makes the difference between an Agent and a workflow clearer. An Agent willObserving(collecting and processing information)Plans(determine what happens next through reasoning),Trading(perform with available tools), andLearning(adjust based on results). This brings me directly to

How an Agent works in the background now

Here you see an example of how to introduce an Agent. A kind of Manager who knows which people are present in his team. And upon receiving a task (in this case, writing an email and performing calculations), he will start by creating a plan to then execute it.

Now, what if we take a technical look at such a solution, what is behind such a solution?

The engine of an Agent is an LLM or large language model. The power of an LLM is thepredictingof words. It actually tries to answer your question as well as possible depending on the text you provide. You may wonder what this has to do with an agent; well, you give the language model your question and also a list ofToolsthat can use it.

ThisToolsyou can see this as functionalities. In the example, the calculator, the (other) language model, and the Email client.

Then the language model will try to estimate as accurately as possible which tools are needed in which order to answer your question, and this will be executed.

For simplicity, I have omitted some more technical terms in the explanation of an Agent, although I would still like to focus on a few.

Terms and Directions that Agents have pursued in the past year

Context engineering

A first term isContext Engineering.Perhaps you are familiar with prompt engineering, where you look to build an optimal prompt. Over the past year, everything has shifted more towards context. The context of a model can be seen as the limit of memory. What is becoming increasingly important for agents is to limit this context. The fewer words used, the lower the cost and the more 'to the point' the answer will remain. (Hence the play on words with context in the intro 😅) For example, it is also easier for people to remember details from 1 book instead of 20 at the same time without mixing them up.

RAG

A next term is RAG or retrieval augmented generation. I don't want to go into too much technical detail here; you can see this as a way to use a language model with the extra limitation that it can only take the answer from various sources. Tools from Google Notebook LM are a good example of this. You can think of it as a kind of selection being made from documents and data to then provide this to the language model.

MCP

MCP of model context protocol is also a term that has come up a lot in the past year. MCP is actually the API or standard 'connector' of tools to language models. This was originally developed by Anthropic (the creators of Claude) and is now also adopted by all the other providers (Google, Grok, OpenAI, etc.). You can see it as the USB-C for connecting functionalities to language models.

SLM

One that did come up sometimes wasSLMsof small language models. These have not received as much attention as their larger counterparts but are quietly improving a lot. This has even reached the point where you can easily run a language model on a laptop without a GPU (graphics card) that still produces understandable output.

These techniques indicate some changes in the field of Agents and will also serve as the basis for further improvements next year.