2025 is anticipated to be the year of Agents (or Agentic AI). In this article, our expert Dr. Julien Siebert explains what agents and multi-agent systems (MAS) are, provides a brief overview of the history of MAS and Agent Oriented Software Engineering (AOSE), compares it with current trends, and discusses the use cases where agents are beneficial.
Sie möchten den Artikel lieber auf Deutsch lesen? Hier geht es zum Artikel: »Agentic AI – Multi-Agenten-Systeme im Zeitalter generativer KI«
Building applications based on generative AI models requires first a profound understanding of each model’s capabilities and how to effectively utilize them in real-world scenarios. This involves knowing what each model can achieve and how to interact with them (i.e., prompt engineering).
However, it is also essential to recognize that models are not standalone solutions but rather key components within a larger architecture. While these models play a crucial role, they are just one part of the ecosystem that enables effective problem-solving.
To harness the full potential of generative AI applications, it is vital to orchestrate the right components tailored for specific use cases. For example, coupling databases with large language models (LLMs) in retrieval-augmented generation (RAG) systems.
One approach to structuring generative AI applications is to divide them into distinct entities known as agents. Each agent is usually assigned specific roles and tasks, allowing them to handle particular aspects of the problem. By enabling these agents to work as a team (and potentially to self-organize), it becomes easier to tackle complex issues.
Which AI applications are right for your business?
Are you interested in leveraging AI and want to know if it makes sense for your business? With »AI Innovation Labs«, we have developed a set of methods and tools that enable us to systematically identify, implement, and evaluate AI applications that offer clear business value.
More about LLMs and generative AI:
German:
- Generative KI in der Softwareentwicklung: Herausforderungen und Szenarien
- Retrieval Augmented Generation (RAG): Chatten mit den eigenen Daten
- Prompt Engineering: Wie man mit großen Sprachmodellen kommuniziert
- Open Source Large Language Models selbst betreiben
- Halluzinationen von generativer KI und großen Sprachmodellen (LLMs)
What is an agent anyway?
A basic definition of „agent“ is given by Stuart Russell and Peter Norvig in their book „Artificial Intelligence. A Modern Approach“ (4th Edition) [Russel & Norvig 21]: „An agent is anything that can be viewed as perceiving its environment through sensors and acting upon that environment through actuators“. Unfortunately, this definition encompasses almost every software system one can think of.
A slightly more precise definition was given by Jacques Ferber [Ferber 99]:
An agent is a physical or virtual entity:
- Capable of perceiving (often partially) and acting in an environment
- which can communicate directly or indirectly with other agents
- driven by a set of tendencies (in the form of an individual goal or a satisfaction/survival function that it tries to optimize)
- which has its resources
- which has only a partial representation of this environment (and perhaps none at all)
- which has skills and can provide services
- which may be able to reproduce itself
- whose behavior tends to fulfill its goal, considering the resources and abilities available to it and depending on its perception, its representation, and the communication it receives.
There are different types (or architecture) of agents depending on how complex their internal reasoning and memory capabilities are. One usually discusses reflex agents for simple rule-based agents and cognitive agents for more complex ones (for instance, agents using the Believe-Desire-Intention architecture).
What is a multiagent system?
A multi-agent system (MAS) is a system in which multiple agents interact with and through their environment. To describe a multi-agent system, the vowels AEIOU [Demazeau 97, Da Silva 02] can be used as a mnemonic: A MAS consists of agents („A“), an environment („E“), agents can interact („I“) with each other and with their environment, they can organize themselves or be part of a predefined organization („O“), which means that organizational strategies or organizational artifacts (such as groups, norms) exist, and often users („U“) can interact with the system.
What is it with the current agentic AI hype?
Multi-agent systems have been researched since the 1990s [Ferber 99, Wooldridge 09], but with the explosion of LLMs, multi-agent systems are being rediscovered by the generative AI community. The perception of agents as a central concept in AI applications of the future can be attributed to a few key factors. First, breaking down complex problems into simpler tasks handled by individual agents proves to be more effective. This should not be surprising to those in software engineering, but this has been empirically demonstrated with generative AI agents for different use cases [Hong 23, Qian 24]. Second, generative AI models are improving in their ability to handle unstructured data (test, images, videos), to think before answering (i.e., Chain of Thoughts), to break down tasks and come up with plans, and lastly, to call functions (also known as tool learning), enabling developers to easily create flexible agents.
How do we build a multiagent system?
- „Can my problem be broken down in different steps?“
- „Do I need separated entities to solve each problem separately?“
- „Are these entities autonomous in some sense?“
- „How will these entities interact to solve my problem?“
- „Do I need to dynamically add or remove entities?“
If you answered yes to most of these questions, then an agent approach is probably a practical candidate.
References
[Demazeau 97] Demazeau, Yves. „Steps towards multi-agent oriented programming“. In 1st International Workshop on Multi-Agent Systems (IWMAS’97), Boston, 1997
[Ferber 99] Ferber, Jacques. „Multi-Agent System: An Introduction to Distributed Artificial Intelligence.“ Addison Wesley Longman, 1999.
[Da Silva 02] Da Silva, Joao Luis T., and Yves Demazeau. „Vowels co-ordination model.“ Proceedings of the first international joint conference on Autonomous agents and multiagent systems: part 3. 2002.
[Wooldridge 09] Wooldridge, Michael. „An introduction to multiagent systems“. John wiley & sons, 2009.
[Russel & Norvig 21] Russel, Stuart and Norvig, Peter. „Artificial Intelligence. A Modern Approach (4th Edition).“ Pearson, 2021.
[Hong 23] Hong, Sirui, et al. „Metagpt: Meta programming for multi-agent collaborative framework.“ arXiv preprint arXiv:2308.00352, 2023.
[Qian 24] Qian, Chen, et al. „Chatdev: Communicative agents for software development.“ Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2024.