Navigating the AI Frontier 2024

Network architecture Supervised architecture AI agentAI agent superivsor AI agent systemAI agent systemAI agentAI agent AI agent systemAI agent systemThe future of AI agents: Towards multi-agent systems2.5 Multi-agent systems (MAS) consist of multiple independent AI agents as well as AI agent systems that collaborate, compete or negotiate to achieve collective tasks and goals.25 These agents can be autonomous entities, such as software programs or robots, each typically specialized with its own set of capabilities, knowledge and decision-making processes. This allows agents to perform tasks in parallel, communicate with one another and adapt to changes in complex environments. The architecture of a MAS is determined by the desired outcomes and the goals of each participating agent or system. There are several architectural types,26 for example: –Network architecture: In this set-up, all agents or systems can communicate with one another to reach a consensus that aligns with the MAS’s objectives. For example, when autonomous vehicles (AVs) park in a tight space, they communicate to avoid collision. In this case, the MAS objective to prevent accidents aligns with each AV’s goal of safe navigation, allowing them to coordinate effectively and reach consensus. –Supervised architecture: In this model, a “supervisor” agent coordinates interactions among other agents. It is useful when agents’ goals diverge, and consensus may be unattainable. The supervisor can mediate and prioritize the MAS’s objectives while considering each agent’s unique goals, thereby finding a compromise. An example could be when a buyer and seller agent cannot reach agreement on a transaction, which is then mediated by an AI agent supervisor. Examples of MAS architecture FIGURE 4: While current efforts largely focus on developing AI agents within closed environments or specific software ecosystems, the future is likely to see multiple agents collaborating in different domains and applications. In MAS, different types of agent could work together to tackle increasingly complex tasks that require multistep processes, integrating expertise from various fields to achieve more sophisticated outcomes. These agents can communicate and interact within a broader adaptive system, enabling them to handle both specific tasks and complex situations more efficiently than a single agent, or even an AI agent system, could on its own. In some cases, multi-agent systems address the limitations of single-agent systems, such as scalability issues, lack of resilience in the event of Source: World Economic Forum Navigating the AI Frontier: A Primer on the Evolution and Impact of AI Agents 14