At Rise of AI 2025 in Berlin, Christian Guttmann, Executive Director of the Nordic AI Institute, argues that large language models are a narrow slice of AI and that the deeper, decades-old field of multi-agent systems deserves more attention as agent deployment accelerates. Drawing on his PhD work on the RoboCup grand challenge, which aims for autonomous humanoid robots to beat the reigning human soccer World Cup champions by 2050, he explains why multi-agent problems differ from static games like chess or Go. He walks through applications he has built, including BDI adherence agents for chronic disease care in Australia, traffic-light coordination researched by Anna Bazzan that Google later showed cut New York traffic by 30 percent and pollution by 10 percent, and micro-simulations of the Swedish population used to model pandemics twenty years before COVID-19. He defines multi-agent systems through cooperating or competing agents with private world models, collaborative planning and game theory such as Nash equilibrium, then lists dos and don'ts for enterprise deployment, warning that naive LLM-wrapped agents and unnecessary autonomy for homogeneous systems are common failure modes. He closes on quantum computing's potential for multi-agent reinforcement learning, followed by a Q&A on investment timing.
Key takeaways
- The RoboCup grand challenge, launched in 1997, aims for autonomous humanoid robots to beat the reigning human soccer World Cup champions by 2050.
- Guttmann's traffic-light coordination research with Anna Bazzan was later shown by Google to cut traffic by 30 percent and pollution by 10 percent in a tested New York area.
- Around 2000, Guttmann built micro-simulations representing every Swedish citizen as an agent to model pandemic spread, work predating COVID-19 by roughly two decades.
- Guttmann warns that deploying LLM-wrapped agents without task allocation protocols or agent modeling is a common cause of failed enterprise multi-agent projects.
- Guttmann argues Europe invests roughly ten times less in quantum computing than the US and says now is the time to take on more investment risk.
Detailed notes
From single-agent hype to the multi-agent grand challenge
Guttmann opens by distinguishing the current wave of large language models, which he calls narrow word models, from multi-agent systems, the subject of his PhD. He introduces the RoboCup grand challenge, launched in 1997 by Hiroaki Kitano and a US-based co-founder, with the goal that by the middle of the 21st century a team of fully autonomous humanoid robots will beat the reigning FIFA World Cup champions under official rules. He argues this domain is fundamentally harder than chess or Go because it requires agents to act in a dynamic, real-time physical world, each with its own perception, and to understand both teamwork and opposition, which single-agent approaches cannot handle.
Applications built over 25 years
Guttmann surveys projects spanning his career. In Australia he built BDI (belief-desire-intention) adherence agents to help patients follow complex chronic disease care plans and support British Telecom customer lifecycle management, using an agent architecture also used by NASA to plan space shuttle missions in the late 1980s. With Anna Bazzan in Brazil he worked on reinforcement-learning agents that coordinate traffic lights to optimize flow and reduce pollution, an approach Google later reported had cut traffic by 30 percent and pollution by 10 percent in a tested area of New York. In Sweden, with the Karolinska Institute, he built micro-simulations representing every Swedish citizen as an agent to model pandemic spread and vaccine rollout around 2000, work later extended to intentional decision-making agents for health authorities and to EV charging demand. He also cites algorithmic trading as a competitive multi-agent domain.
What actually defines a multi-agent system
Guttmann defines a multi-agent system as a collection of cooperating and competing agents with joint or individual goals, each holding a private assessment of the world inaccessible to other agents. He stresses agents need the ability to plan, ideally collaboratively, and to build models of other agents' capabilities and intentions, citing research on shared plans and on flexible teamwork that accounts for communication failures or team members dropping out. He highlights game-theoretic concepts such as Pareto optimality and Nash equilibrium as essential when agents operate in adversarial or competitive settings, and mentions recursive modeling, where agents repeatedly model each other's likely strategies, as one method for learning to anticipate other agents.
Dos and don'ts for enterprise deployment
Guttmann warns that simply wrapping an LLM with a name and role and deploying it as an autonomous agent is a common failure mode in industry. He recommends established multi-agent foundations rarely used in practice, including contract net protocols for task allocation, formal agent modeling, and decentralized planning. He cautions against overkill: giving multiple agents full autonomy when they share the same data and control is unnecessary complexity, and rigid master-slave architectures misjudge what makes something an agent rather than a passive object. He also stresses the open system assumption, that real deployments involve agents with bounded rationality and possible failures, unlike simplified enterprise assumptions of full visibility and reliability.
Q&A highlights: quantum computing and European investment
Guttmann closes by describing a quantum computing spin-off from the Nordic AI Institute focused on pharma applications, calling quantum computing promising for multi-agent reinforcement learning problems. An audience member identifying as a partner at AI Fund and a quantum physicist pushes back, arguing that with only about 50 usable qubits and short decoherence times, practical value will not arrive until qubit counts and coherence reach roughly 10,000, with exceptions for quantum simulation and certain cryptographic problems. Guttmann responds that Europe already invests roughly ten times less in quantum computing than the US and argues this is exactly the moment to take on more risk rather than repeat what he frames as Europe's earlier underinvestment in AI.
Quotes
Building naive agents and have them run at scale, just an LLM wrapper, is doomed to fail.
Christian Guttmann
If agents have access to the same data and they have the same control and you start building a system where you give several entities autonomy, is a complete overkill.
Christian Guttmann
We have already ten times less investment in quantum here in Europe as we have in the US, so let's not miss that train.
Christian Guttmann
Questions answered in this talk
What is the RoboCup grand challenge in AI?
According to Christian Guttmann at Rise of AI 2025, RoboCup was launched in 1997 with the goal that by the middle of the 21st century a team of fully autonomous humanoid robots will beat the reigning FIFA World Cup champions under official rules, requiring agents to perceive, cooperate and compete in a dynamic real-time environment rather than a static game like chess.
What is a common mistake companies make when deploying multi-agent AI systems?
Christian Guttmann says at Rise of AI 2025 that simply wrapping a large language model with a name and role and deploying it as an autonomous agent, without task allocation protocols or agent modeling, is almost doomed to fail, and that giving multiple agents full autonomy when they share the same data and control is unnecessary overkill.
Has AI-based traffic light coordination shown measurable results?
According to Christian Guttmann at Rise of AI 2025, reinforcement-learning traffic-light coordination research, building on work by Anna Bazzan, was later applied by Google and reported to reduce traffic by 30 percent and pollution by 10 percent in a tested area of New York.