Trend Topic:  Agentic AI

Agentic AI is the new buzzword in automation — an LLM (Large Language Model) that acts autonomously, makes decisions, and learns from experience. In the following sections, you will find out everything you need to know about Agentic AI.

LLMs control their own processes

On the one hand, there are agents in which LLMs dynamically manage their own processes and the use of tools while maintaining control over how they execute tasks.
Example: A drone that independently flies from point A to B in any weather conditions and is capable of avoiding obstacles.

LLMs with predefined code paths

On the other hand, there are agents in which LLMs and tools are orchestrated through predefined code paths.

Example: Case registration in a hospital.
A doctor sends a referral letter via email to the hospital. There, Agentic AI extracts the context and the structured information needed for case registration from the letter. The bot can then automatically record the case in the hospital’s system.

A Large Language Model (LLM) is a deep learning algorithm trained on enormously large datasets. LLMs are often used in applications within the field of Natural Language Processing (NLP) when it comes to entering queries in natural language to receive an answer or result. An LLM can understand, summarize, generate, and predict new content. It typically has billions of parameters that function like memories the model accumulates during training. This enables it to derive new content.
The most well-known LLMs include GPT from OpenAI, Claude from Anthropic, and Llama from Meta, to name just a few.

This component retrieves relevant information from external sources or databases. It ensures that the agent always has up-to-date and relevant data to make decisions.

With assigned tools, the agent can perform specific tasks, such as calculations, executing programs, or using APIs. The tools expand the agent’s capabilities by helping it solve problems or providing additional functions.

Agentic memory stores information that the agent collects over time, such as previous interactions or results. This allows the agent to draw on past experiences to improve its decisions and better understand the context.

Today, almost exclusively, agents with predefined code paths are used in companies and organizations. The reason for this is obvious. The most crucial component of an agent is the LLM (Large Language Model)—a pre-trained model with billions of parameters, capable of generating new content.

However, LLMs only ever predict the next token (letter, word) based on probability. They do not understand the context of a sentence in the way humans do. This is why it is essential to implement automated validation processes (code paths) in the background to detect and correct hallucinations in LLM outputs.

No one in the business world wants to miss out on the capabilities of LLMs (such as GPT). But how can this be reconciled with data protection, especially when dealing with sensitive data? One way to ensure complete security is to use an on-premises (offline) LLM server that operates without internet access.

For the past few weeks, we’ve had our own NVIDIA GPU-based server—and we’re thrilled by how well offline LLM applications can be implemented. So, what does our server look like, and where is it located?

When LLM-based agents are used in automation solutions, it is almost 100% certain that a validation component will also be required. This ensures that missing or incorrect information can be corrected or supplemented by the user. Validation apps should always be “built for purpose” with a modern and intuitive user experience.

Various providers, such as UiPath, Automation Anywhere, or Microsoft, offer frameworks for building agents. These low-code/no-code platforms allow agents to be developed quickly for specific tasks.

In contrast, the API approach brings agents to life through a software development process. Depending on the task, complexity, licensing costs, and other factors, one approach may be more suitable than the other. However, one thing remains the same: a solid understanding of software technology is essential to successfully implement and operate processes using this exciting technology.