Numerous programming libraries and services for developing multi-agent systems have been established in recent years. These include CrewAI, LangGraph and more.
But what are multi-agent systems?
Several AI agents together form a multi-agent system. Each AI agent possesses individual characteristics such as specific goals, action plans, underlying large language models (LLMs), and tools, enabling them to solve tasks collaboratively. Possible tools for an AI agent include, for example, interactive web search, database access, and access to information retrieved via RAG (Retrieval-Augmented Generation). Overall, this system can be compared, for instance, to a virtual IT company where different employees with various roles (software developer, software tester, manager, CEO) collaborate to solve tasks (see: https://github.com/OpenBMB/ChatDev?tab=readme-ov-file).
What benefits can we derive from this?
Even today, multi-agent systems – given the current state of technology – are suitable for repetitive and simple tasks. Potential tasks for such a system include, for example, automatic communication via email, marketing tasks, customer acquisition, data mining, data cleaning, market and data analyses, remote sensing, and even smaller software development projects. Connecting numerous multi-agent systems to form more complex overall systems is, of course, also possible. However, for the final review of the results, it is still advisable to involve humans to ensure the quality of the outcomes (see: https://www.answer.ai/posts/2025-01-08-devin.html). But with the increasing improvement in the quality of language models, the output results of multi-agent systems also become more precise, which could lead to more autonomous systems in the near future.
From a developmental perspective, multi-agent systems can be classified as an advanced stage on the path towards Artificial General Intelligence (AGI) – potentially an essential step towards the complete automation of companies or institutions. However, it is also becoming apparent that large language models – currently often used as the fundamental building blocks for multi-agent systems – cannot yet comprehensively model human intelligence and the physical world on their own (see Yann LeCun: Lecture Series in AI: “How Could Machines Reach Human-Level Intelligence?”: https://www.youtube.com/watch?v=xL6Y0dpXEwc and Yann LeCun: Why Can’t AI Make Its Own Discoveries? https://pod.link/1522960417) and are therefore likely only an intermediate step towards AGI. Perhaps physics-based world models like V-JEPA (see https://ai.meta.com/blog/v-jepa-yann-lecun-ai-model-video-joint-embedding-predictive-architecture/) in combination with large language models (LLMs) represent another step that could also improve the precision of multi-agent systems and represent the next step towards AGI.
Nevertheless, we are still far from AGI. The timeframe for AGI’s arrival is estimated differently by scientists: According to Demis Hassabis, it could arrive in five to ten years (see https://www.cnbc.com/2025/03/17/human-level-ai-will-be-here-in-5-to-10-years-deepmind-ceo-says.html), Geoffrey Hinton suggests a timeframe between five and twenty years (see https://www.youtube.com/watch?v=MGJpR591oaM), and Yann LeCun mentions that AI could reach human-level intelligence in ten years in an optimistic scenario (see https://www.youtube.com/watch?v=JAgHUDhaTU0).
Therefore, it is crucial to consider and mitigate the risks associated with current and future AI development on the path to the AGI era. Mustafa Suleyman has addressed this in detail in his book ‘The Coming Wave’, and the European Union has also established a legal foundation with the EU AI Act (AI Regulation) to protect against risky AI developments. It is crucial to shape development in such a way that possible dystopian scenarios, such as those described by Max Tegmark in ‘Life 3.0’, are avoided. The aim must be to ensure that AI contributes to a desirable life for future generations.
References:
https://github.com/OpenBMB/ChatDev?tab=readme-ov-file
https://www.answer.ai/posts/2025-01-08-devin.html
https://www.youtube.com/watch?v=xL6Y0dpXEwc
https://pod.link/1522960417
https://ai.meta.com/blog/v-jepa-yann-lecun-ai-model-video-joint-embedding-predictive-architecture/
https://www.cnbc.com/2025/03/17/human-level-ai-will-be-here-in-5-to-10-years-deepmind-ceo-says.html
https://www.youtube.com/watch?v=MGJpR591oaM
https://www.youtube.com/watch?v=JAgHUDhaTU0
