After the introduction of ChatGPT, many generative AI applications have adopted the Retrieval Augmented Generation (RAG) pattern, focusing on the variation of a chat over a collection of documents. Currently, the focus is to make RAG systems more robust and shape the next generation of AI applications where common themes are centralized. These agents are designed in a way that helps the Language Model (LM) to enhance its capabilities in solving real-world problems. The main requirement of agents to solve real-world problems efficiently is their ability to reason, plan, and execute tools efficiently.
AI agents have become an important part of generative AI applications. However, for effective interaction with complex environments, they need a strong ability to reason to make independent decisions and help users solve various tasks. The tight synergy bond between acting and reasoning is very helpful for AI agents to learn new tasks quickly. Moreover, AI agents require reasoning to modify their plans, considering new feedback or information learned. Lack of reasoning skills may lead to these agents’ improper working, including misinterpretation of the user’s query, failure to consider multi-step implications, etc. All these shortcomings of AI agents have been addressed in this paper.
A team of researchers from IBM and Microsoft introduced AI agent architectures to achieve complex goals, including enhanced reasoning, planning, and tool execution capabilities. It consists of (a) Single Agent Architectures (SSAs) and (b) Multi-Agent Architectures (MAAs). These two architectures help recognize key patterns, divergences in design choices, and overall impact evaluation on achieving a particular task. SSAs are supported by one language model and perform all the reasoning, planning, and tool execution independently. In contrast, MAAs consist of two or more agents and each agent leverages the same language model or a set of different language models.
Single Agent Architectures: In this type of agent, there is no concept of a feedback mechanism from other AI agents, but there is an option of feedback provided by users that helps guide the agent in achieving its goal. SAAs exhibit good performance when their characteristics and set of tools are defined.                                                      Multi-Agent Architectures: These types of agents have their characteristics or identities, and they have the accessibility to use the same tools or different tools. MAAs contain multiple organizations, which are divided into vertical and horizontal fractions, indicating the two ends of a spectrum. Most of the architecture lies between these two boundaries.
While working on SAAs and MAAs, researchers introduced some methods, including a specific stage for analyzing the problem before taking action to achieve the goal. Language Agent Tree Search (LATS) is a single-agent method combining planning, acting, and reasoning using trees. This technique uses an LM-based strategy to search for possible outcomes and then uses a state evaluator to select an action. Another method, MetaGPT, is used to overcome the challenge of MAAs, which is the unproductive chatter amongst agents by requiring agents to generate structured outputs like documents and diagrams. When compared with SAAs on the HumanEval and MBPP benchmarks, MetaGPT’s MMA exhibits better results.
In conclusion, Researchers proposed AI agent architectures to achieve complex goals, including enhanced reasoning, planning, and tool execution capabilities. It consists of (a) Single Agent Architectures (SSAs) and (b) Multi-Agent Architectures (MAAs). Both single and multi-agent patterns perform strongly on various complex tasks involving reasoning and tool execution. Besides, these architectures face challenges with agent evaluation. For example, many research teams introduce their unique agent benchmarks alongside their agent implementation, which makes comparing multiple agent implementations on the same benchmark challenging.
Check out the Paper. All credit for this research goes to the researchers of this project. Also, don’t forget to follow us on Twitter. Join our Telegram Channel, Discord Channel, and LinkedIn Group.
If you like our work, you will love our newsletter..
Don’t Forget to join our 40k+ ML SubReddit
The post Single Agent Architectures (SSAs) and Multi-Agent Architectures (MAAs): Achieving Complex Goals, Including Enhanced Reasoning, Planning, and Tool Execution Capabilities appeared first on MarkTechPost.
Source: Read MoreÂ
