AI agents are coming

The authors of this article, Filip Tichý (Partner at Grant Thornton Slovakia) and Jakub Chudík (Co-Founder at Assetario), take you through the world of artificial intelligence in the AI Breakfast series. This article was written without the use of AI.

At the end of last year, Henry Kissinger's last book, Genesis, co-written with Eric Schmidt and Craig Mundie, was posthumously published. This was his second book on AI, and unfortunately also the last. The main topic of the book is what will it mean to be a human in the age of AI. However, the book also partially deals with the latest AI trend – the so-called „AI agents“. In the coming months and years, after the LLM revolution triggered by Chat GPT, the most prominent manifestation of AI will be the "Agentization". AI agents technology is not some new app or new technology. They represent a new, expanded way of using AI models for an extended process, and are another significant step in the gradual implementation of AI into the everyday lives of people and companies. While LLM applications, such as Chat GPT or Gemini, operate on constant interaction via prompts, agents can independently perform more complex tasks without the need for constant human interaction. They are based on the stacking of multiple multimodal LLMs and foundational models, that are combined into a coherent AI workflow designed for a specific process. Do you want a new house? An AI agent will be able to do the research, design a project that suits your needs and financial capabilities, secure permits, procure purchases, pay invoices and, if necessary, sue your builders over construction delays or budget overruns. Also, a simpler form of agents is beginning to emerge. Handling customer complaints, sending out personalized offers, preparing analyses and handling emails will be gradually transitioned to AI agents over the next two years.

How does it technically work?

An AI agent is actually an AI workflow, that is designed to effectively execute a given process, and to be suitable for an AI agent. That means, flipping a corporate business process over to an AI agent will never be the same 1:1. The agent workflow is coded in a programming language (e.g. Python) or "assembled" in a low-code/no-code workflow platform (e.g. Trilex AI, n8n, Flowise). Not only does the AI agent have their workflow set up, but it also gets access to applications it need to use (calendar - so the agent can check availability and schedule events, credit card - so the agent can place orders, booking.com - so the agent can research availability or bookings, Chat GPT - so the agent can write emails, ERP system - so the agent can record new transactions to the system, etc.). The role of the agent is also to translate the natural language into the format of the tool it is going to use. For example, to search for accommodation on Booking.com, you need the date of travel, location, budget per night, room amenities, etc. If this information is not available after the first prompt, the agent tries to get it by subsequent queries or from the context of the user's gmail or outlook calendar, so that it can interact effectively with the tool. Unlike RPA workflows, AI agent can solve even ambiguous situations, autonomously make smart decisions and perform complex tasks that require working with different types of inputs and outputs. However, the AI agent still operates on the basis of a predefined workflow. It doesn't begin actions it doesn't have predefined. At the end of the agent's process, a "review loop" can be set up to independently check the proposed output, even a hundred times over and over. This allows for AI agents to make more complex decisions and outputs.

After the workflow is programmed, what follows is a phase of debugging, testing and validation of outputs, efficiency checking and finetuning. What else is important when designing AI agents:

• To have a human input in the agent's cycle at appropriate points (e.g. at the beginning and before the execution of binding and important actions).
• Appropriate technical design, the individual applications that are part of the agent workflow must be compatible with each other (for example via MCP server).
• To keep transparency and explainability (the work and results of an AI agent's work should be transparent, explainable and clear to humans and other colleagues).
• The reliability of an AI agent can be a "black box", it is important to be able to rely on the agent to produce 100 times consistent output when executing 100 iterations.
• Scalability and flexibility (it is crucial for a successful implementation, you can invent an AI agent application for practically anything, but it makes sense economically mainly if the usecase is scalable).

AI agent = new colleague?

The task for an AI agent can be simple and short, or complicated, long, containing multiple levels and multiple agents, who also can have an AI agent manager.  AI agents are already on the verge of being seen as an actual team members. The potential for the use of AI agents in companies is huge. The race about how to improve productivity, performance and automate as much as possible (as well as marketing hype) is on.