Multi-Agent Systems
Multi-Agent Systems in Agentic Reasoning involve multiple models working together to execute Reflection, Planning, and Tool Use in a collaborative or competitive manner. These agents can communicate with one another, exchange feedback, and coordinate tasks to solve complex problems. Tends to improve upon Single-Agent Systems more when collaboration and multiple distinct execution paths are required 1.
In this system, agents may have specialised roles and communicate with one another to enhance performance, such as through division of labor or consensus-building. While human guidance can still be integrated, the focus is on agents autonomously interacting to optimise the outcomes.
According to 1, at the two extremes, there are two main categories of multi-agent system pattern.
Vertical Architectures
In this structure, one agent acts as a leader and has other agents report directly to them. * Depending on the architecture, reporting agents may communicate exclusively with the lead agent. * Alternatively, a leader may be defined with a shared conversation between all agents. * The defining features of vertical architectures include having a lead agent and a clear division of labor between the collaborating agents.
Horizontal Architectures
In this structure, all the agents are treated as equals and are part of one group discussion about the task. * Communication between agents occurs in a shared thread where each agent can see all messages from the others. * Agents also can volunteer to complete certain tasks or call tools, meaning they do not need to be assigned by a leading agent. * Horizontal architectures are generally used for tasks where collaboration, feedback and group discussion are key to the overall success of the task.
Key Papers
The Landscape of Emerging AI Agent Architectures for Reasoning, Planning, and Tool Calling: A Survey * Surveys existing single and multi-agent architectures, and defines the Vertical and Horizontal architectural patterns.