Multi-agent systems; a Society of Minds or just the Worst Unsustainable Way of doing things?
For the better or worse, Generative AI has flabbergasted us, but also made the shortcomings of brute-force AI very visible to the masses. But it did that with a Bang! Never have we collectively been so impressed, scared or something inbetween, by a machine that interacts with us. Or did it actually interact with us? It seems that much of Generative AI is reactive, it needs to be prompted, kicked off, initiated. And even then there are many tasks it cannot perform or struggles to get it right. We have looked for, and found, ways around some of the issues and learned how to “play the piano”. But more often than not the results are non-deterministic, require difficult to repeate work-arounds, stopping to work when going to the next model or version. And some things are simply not correct(able). Causing you to wonder “why can´t we fix this”? Scaling is one of the proposed solutions for inaccuracies and lack of quality. Some even promote to fully dedicated (all the US/world´s?) resources to create something, a monolite of sorts, that can “do it all”. If we don´t do that, we/us/them might be doomed (some people think so at least: see ref 1). With a capability improvement that seems to plateau, and no clear path forwards to significantly enhance trustworthiness and correctness with Generative AI, other routes might be investigated.
Distributed Intelligence
Except for in religion, there seem to be no singular, benign and superior organisms on this planet. The intelligence that exists only works because it is distributed over many individuals. Be it delphins, octopi (2), humans, primates or other organisms, their intelligence is divided over many brains. Each and every of those brain add to the equation and has specific capabilities or deviations. Social interaction, co-existence, together the results are enough to lead to evolving capabilities, innovation and, in some cases, self-awareness, abstraction, reasoning, and planning capabilities. Finally such collaborative work might result in teams being able to build tools, make fire, build rockets, design medicine, computers, create social systems, political systems, logical frameworks, math, physics, chemistry and you name it. No single mind is able to develop or use all of it. The last quintessential polymath of the world could well have been Leonardo da Vinci for his wide mastering of sciences. As human creations and their interactions with the natural and virtual worlds become more complex, it’s unrealistic to expect a single machine to handle all that complexity effectively. Or even unwanted. Because it would have to be able to anticipate and deal with many contexts, cultures, ideas, needs, necessities of very many organisms on our planet. Because the current tech can learn but cannot forget. It can regurgitate, but not abstract and model real-world contexts.
So what might be a solution? Let´s follow a path of reasoning that Marvin Minsky did in his book “Society of Minds”. In the context of AI, Minsky’s theory suggests that to create intelligent machines, we should design them as a collection of smaller, specialized programs that can collaborate. Each program would handle specific tasks, and their interaction would result in intelligent behavior. This approach contrasts with the idea of building a single, all-encompassing AI. Instead, it’s about creating a network of cooperating modules, each contributing to the system’s overall intelligence.
“A Multi Agent System is a system of collaborative actors that work in a coordinated way to solve tasks of a variety of complexities”
Nice definition, but this can be built in very many different ways, with very different capabilities and characteristics if it comes to capability, robustness, flexibility, sustainability, maintainability and quality.
Agentic systems using AI- agents
But let´s start simple, a single agent helping you as the (long perceived?) personal assistent. It can get your orders, follow your calendar, help you with whatever questions and can query “the internet” for specifics if needed. Such personal digital agents (PDA) able to interact with you in natural language have existed for a decade now (Alexa, 2013, and simpler PDAs before that). Based on the current generation of Large Language Models (LLM), such PDAs are starting to perform impressively well. Their (multi-lingual) capabilities are unequaled and also the way they can follow a discourse is impressive. At the current multi-modality, where your assistant can interpret images, sounds and combine it with text. Based on foundation models, they tackle a wide variety of topics, but with varying levels of success. And it is not necessarily easy to find out when it fails. To that end AI Agents can be trained and triggered for specific tasks, but not all tasks. This is what we see in current solutions, and market places for AI Agents aimed at specific domains, tasks or routines are growing.
Swarm Intelligence — many of the same
An early form of multi-agent systems (MAS) was very focussed on robustness, resilience and survival in hostile environments. They where made up of agents with the same capability, cheap and many. You might remember Agents chasing Mr Anderson in the kult movie “Matrix”? In the real-world situations where multi-agents with such swarm intelligence are used are typically monitoring and search & rescue situations. For example sending out many drones in all directions, covering large areas with redundancy allowing for fall out of a certain number of such robots, without effect on overall performance. But you can also imagine other hostile environments, like forrest fires, underwater pipeline inspection, mine-fields, underground rescue, you name it. Characteristical for these scenario´s is that they are multi-agent, cheap and redundant. Such systems are typically autonomous, with all entities pursueing the same goal. A robocup team with such swarm intelligence agents might cluther around the ball all together. The approach showed to be too naïve in various scenario´s and so a bit more differentiation and coordination had to be added.
Polymath systems — many with specialised capabilities
Taking that to the other end of the scale on multi-agent sytems layouts, that show a lot of variety in capabilities of each single agent. A bit like in an engineering team, you can have many different roles and competencies needed to build and use a complex product. Such a Multi-Agent system might have a material expert, a chemistry agent, an agent great in designing of physical products with specific characteristics, and why not a regulatory agent, an ethical agent, one that keeps an eye on the supply chain and so on. Each and every one adds to the equation. And the best with it? Such systems can flexibly configure solutions for specific tasks, and if not needed or obsolete, such a system can easily “forget”, but taking out agents. Commercially such systems might also fit much better to our current society and economy, as they allow for organisations and individuals to build their own agents, based on their own knowledge and data. Offering those to third parties might make an agent economy with incentives fitting well to current economic systems. And it will keep each data and capability owner in control of their own competitive advantage. So no need to give away your data, knowledge and capability. Trade it, or share, but decide for your self.
And with this, another interesting dimension hits the fan. Some multi-agent systems concists of agents converging to a common goal, they want to cooperate and build upon each others results. But in other scenarios (e.g. a Dutch Auction scenario), agents are competing. In both cases being able to interpret the intention of the agents you interact with is of utmost importance. And it has to induce very different behaviour depending on which context the agent operates in. As we have seen, many current AI implementations struggle with the correct interpretation of their execution in the real world. Multi-agent systems might bring this struggle to the virtual world. But that might be just the opportunity we need to be able to find solutions for context interpretation.
Many colours, many tasts
An interesting thought experiment is to have a look at what types of capabilities agents can show. Indeed, can we identify axes along which we could classify agents? This might indeed become necessary if we have to evaluate performance, do risk assessments, or even find out about costs vs needs in order to prevent over-engineering of solutions.
- Naïve Task-Executors: These agents perform tasks without needing a reason, similar to basic tools or appliances. Tasks are pretty straight forward and often pre-defined, as in a web-API returning the next-day weather prediction for a specific spot on the planet.
- Capability Agents: Agents using pre-defined models which estimate or describe parts of the real world. This can be mathematical, physical or even logical models. They represent a slightly more complex type of task-executing agents. For example, an agent might assist with mathematical algorithms to optimize logistics in a warehouse or use a physics model to calculate the strength of a construction.
- Goal-driven Agents: These agents perform tasks because they have a goal, like a robot that cleans because its goal is to keep the environment tidy. Goals can be single (non-decomposable goals), but more often than not, goals are a composition of sub-goals to execute, and planning for their order of execution becomes a part of the work.
- Social goal-driven Agents: The most advanced agents, capable of communication, negotiation, and cooperation with other agents to achieve complex goals. These agents are like virtual assistants that can plan and execute tasks orchestrating other agents as necessary.
As you might appreciate, these agents can be divided into “reactive” vs “goal-driven”, akin to the “System 1” and “System 2” thinking coined by Daniel Kahneman (3). The first two types of agents are merely executing without the need for negotiation or social capabilities. Goal-driven agents could, according to their goals, initiate independent actions. The last agent type adds communication and negotiation skills to the game. This might even include the concept of being able to capture “intent” (e.g of the other parties in a game) and building strategies for maximizing goal fulfillment in a competitive scenario.
Sustainability in a world in turmoil
MAS require significant computational resources, especially when handling complex tasks across multiple agents. For all these different scenario´s, the need for compute (and thus cost and sustainability) varies heavily. MAS can run with and without AI, and when using AI there are huge differences in efficiency of the various models, their sizes, their level of adaptation (through fine-tuning or similar) and their implementations on various types of hardware. Some models are available for local execution with a small-footprint, others run online through services paid with tokens for compute.
Sustainable MAS must also be designed with ethical and societal impacts in mind. As MAS become more integrated into daily life, their design needs to account for fairness, transparency, and the potential societal consequences of their deployment.
The sustainability of multi-agent systems using AI is a growing concern, especially as these systems become more complex and widespread. Sustainability in this context involves both environmental and operational factors.
- Energy Efficiency: MAS require significant computational resources, especially when handling complex tasks across multiple agents. Advances in AI and machine learning, like more efficient algorithms and hardware optimization, might help to reduce the energy consumption of parts of solutions. Developing overall energy-efficient MAS is crucial for minimizing their environmental footprint.
- Scalability and Maintenance: For MAS to be viable in the long term, they need to be scalable and maintainable. This means designing systems that can adapt and evolve without requiring excessive resources or frequent human intervention. Automated processes are helping make MAS more self-supporting by reducing the need for manual tuning and updates. But scaling up can be rather unsustainable if it comes to overall footprint after all.
- Ethical and Societal Considerations: Sustainable MAS must also be designed with ethical and societal impacts in mind. As MAS become more integrated into daily life, their design needs to account for fairness, transparency, and the potential societal consequences of their deployment.
- Resilience and Adaptability: A sustainable MAS must be resilient and robust to changes in its environment and adaptable to new challenges. This includes being able to operate under varying conditions without significant drops in performance, which is essential for long-term viability.
Multi-agent systems represent an exciting field where intelligent agents work together to perform tasks of varying complexity with autonomy. They can be simple or complex, organized in different ways, and have specialized roles. The control of these systems can be centralized or decentralized, depending on the needs, goals and level of autonomy these agents have. Understanding MAS is a prerequisite for a future where AI is well-integrated into our daily lives, assisting in everything from household chores to industrial processes and even space exploration.
(1) Situational Awareness. Leopold Aschenbrenner, june 2024
(2) Children of Ruin. Adrian Tchaikovsky, 2019
(3) Thinking Fast and Slow. Daniel Kahnemann, 2011
