Multi-Agent System’s Architecture

The distribution of decision-making and interaction among the various agents that make up the system principally distinguishes multi-agent systems from single-agent systems. In a single-agent system, a centralized agent makes all decisions, with other agents acting as remote slaves. It is customary for this one agent to decide depending on the circumstances. This can lead to the overlooking of alternative viewpoints and possibilities. On the other hand, multi-agent systems consist of several intelligent agents that interact with each other, each capable of making decisions and influencing the surrounding environment.

The purpose of multi-agent architecture is to construct agents that are able to bring in multiple perspectives by virtue of the roles that they play. Different contexts facilitate the creation of these agents. Despite using the same LLM, each agent’s behavior is unique due to its specific function, objective, and context, just like a squad member.

Imagine that you have an agent who is responsible for generating application code and another agent who is responsible for reviewing the code, and both of them engage in a conversation with each other in order to enhance the code. These two agents will achieve the best outcomes after completing a predetermined number of iterations. The application of this strategy has a significant potential to not only provide a more desirable product but also to lessen the impact of hallucinations, bias, and other similar phenomena. In a multi-agent architecture, defining the appropriate context, prompt, and model is crucial. With strong, quick engineering abilities, it is possible to create an impactful application that utilizes multiple agents.

Multi-Agent Design Advantages

Here are a few advantages of using multi-agent designs:

Separation of Concerns

Different language models, fine-tuned and supported by a variety of tools, drive each agent’s instructions and few-shot examples. This allows for the separation of responsibilities. Delegating work to multiple agents can achieve better outcomes. It is possible for each agent to concentrate on specific tasks instead of having to choose from a wide variety of tools.

Modularity

Multi-agent designs, due to their modularity, enable the division of complex tasks into more manageable segments that specialized agents and language models can target. When you use multi-agent designs, you can evaluate and improve each agent independently without interrupting the program as a whole. Grouping tools and responsibilities is likely to yield better outcomes. Focusing agents on specific tasks increases their likelihood of efficiency.

Diversity

Ensure that the agent teams have a significant amount of diversity in order to incorporate a variety of viewpoints, improve the quality of the output, and stay clear of hallucinations and bias. This is similar to a traditional human team.

Reusability

Once the agents have been developed, there is the possibility of reusing them for a variety of use cases. Additionally, it is wise to think about the possibility of an ecosystem of agents that can collaborate to find a solution to the problem, utilizing an appropriate orchestration framework (for example, AutoGen, Crew.ai, etc.).

Source: Multi-Agent Architecture

Multi-Agent Architecture Components

Following are the components of a typical multi-agent architecture.

• Agents: Intelligent agents operate on an LLM and are characterized by having a clear purpose, persona, and context.
• Connections: What are the links between these agents?
• Orchestration: The term “orchestration” refers to the process by which these agents collaborate with one another (such as sequential, hierarchical, or bi-directional chat).
• Human: We will need to have a human in the middle so that we can assist with making decisions and evaluating the outcomes.
• Tools: These agents use tools to perform specific tasks, like searching the internet for more information, generating and reading documents, uploading generated code to GitHub, and so on.
• LLM: The agent will use large language models to support the design for inference.

Frameworks

Many frameworks facilitate the development of these multi-agent applications. In this section, we will explain some of these frameworks in detail. However, as this sector evolves, numerous frameworks undergo changes and introductions. On the other hand, each of them offers easier techniques for constructing and managing many agents. As we move forward with our blog posts, we will be developing some applications based on these frameworks.

Open AI Assistant

One of the first multi-agent frameworks was OpenAI Assistant. This framework enables the development of multi-agent systems that are persistent, multi-modal, and capable of interacting with users over long periods. Agents can collaborate to complete tasks by accessing files, tools, and a code interpreter. This is suitable for long-term collaboration and interaction applications.

Microsoft Autogen

Microsoft’s Autogen framework is well-known for being one of the most popular emerging frameworks. For constructing strong multi-agent applications, this open-source framework includes Autogen Studio, an intuitive UI-based programming tool. It creates LLM agents that use large language models for reasoning and action, as well as bespoke data. It provides a specified multi-agent architectural method based on an orchestrator.

Dragonscale’s Multi-Agent Systems

Dragonscale’s Multi-Agent Systems integrates generative AI models and tools to develop intelligent systems that can handle simple to complicated tasks, especially in dynamic business situations. This is suitable for business situations that are constantly changing and require agility. The framework assists in handling complicated tasks.

CrewAI

We are comparing the recently popular framework CrewAI to Autogen. CrewAI is a highly useful framework for orchestrating autonomous AI agents. CrewAI gives agents the intelligence to collaborate on challenging tasks. It helps AI agents to adopt roles, share goals, and work together. This framework is not only one of my favorites, but I am closely monitoring its progress. In my next blog post, I will discuss the apps that I have developed with CrewAI.

LangGraph

LangGraph is yet another multi-agent framework that is extremely powerful and promising. Developers build LangGraph on top of LangChain to develop stateful, multi-actor applications using LLMs. It enhances LangChain Expression Language by adding cyclic coordination of numerous chains (or actors) over various processing steps, drawing inspiration from Pregel and Apache Beam. The strength of the LangChain environment and the community that supports it gives LangGraph its power.

Closing Thoughts

Intelligent systems that learn from and adapt to their environments are the ultimate objective of AI evolution, which leads to more personalized and efficient solutions across industries.

Thank you for reading!

Let’s connect on LinkedIn

Sign up to my free newsletter.