Agentic AI Design Patterns Introduction and walkthrough | Amazon Web Services

Key Takeaways

Hi everyone, welcome to the session. Agentic AI is a rapidly evolving space and it will fundamentally reshape how we work across processes, team dynamics and entire industries. I'm Romeshalla joined by my colleague Praep Sridtheran. We are both solutions architects working at AWS and in this session we will learn more about the foundational design patterns for agentic AI systems. Let's start by understanding what agentic AI means and where does it fit in the generative AI landscape. The evolution of generative AI has followed a clear progression with each phase bringing new priorities and challenges for businesses. In the past, AI models used to be trained on specific data sets to solve specific problems. That required a lot more human oversight to see if the AI was getting it right. As we continue to evolve towards a fully realized agentic world, intelligent multi- aent systems will make complex decisions and coordinate effortlessly with less human oversight, completely transforming the way we work and innovate. AI agents are autonomous software systems that have the ability to think iteratively, to evaluate results, adjust their approach, and continue working towards a defined goal. They don't just answer questions. They solve problems through a process of exploration and refinement. At its core, agentic AI refers to AI systems leveraging distributed intelligent agents that not only act independently but also collaborate to solve highly complex real world problems. Agentic AI workflows represent a significant leap forward and a paradigm shift in AI systems design with features including autonomy, adaptability, and decision making. At the core of agentic AI workflows are key design patterns that enable large language models to exhibit more autonomous and intelligent behavior. These key design patterns are reflection, tool use, planning, and multi-agent. Reflection enhances AI's ability to evaluate and improve its own outputs. Through this, AI can provide more accurate and reliable results. The tool use pattern enables AI to expand its capabilities by utilizing external tools and resources. Planning pattern is particularly useful for managing long-term projects and solving complex problems. In multi- aent pattern, multiple AI agents collaborate to perform complex tasks. Each agent takes on specific roles to contribute to achieving the overall objective. Let's now dive deeper into each of these design patterns. Reflection is a technique where AI models self-evaluate and refine their own outputs. This pattern enables AI models to become more autonomous, creative, and reliable by mimicking human-like feedback and revision loops. It is particularly useful for large language models, allowing them to catch mistakes, clarify ambiguities, and improve over multiple iterations. This pattern involves generation where the AI model generates an initial response to a given prompt or task. Self-reflection where the model scrutinizes its output identifying potential errors, inconsistencies or areas for improvement. Iterative refinement where based on the self assessment, the model iteratively refineses its output making adjustments to enhance its quality and accuracy. Let's look at an example implementation of the reflection pattern using Amazon Bedrock. Amazon Bedrock is a fully managed service that offers a choice of high-erforming foundation models, making it easy to build and scale generative AI applications. In this architecture diagram, you can see the two core steps of the reflection pattern, initial generation and self-evaluation. The first AI model solves the initial problem and provides the output and the second AI model validates the solution demonstrating how AI can self-evaluate its results. Moving on to the tool use pattern. This pattern enables large language models to transcend their natural limitations by interacting with external functions to gather information, perform actions, or manipulate data through tool use. Large language models are not just confined to producing text responses from their pre-trained knowledge. They can now access external sources and functions to gather latest information, process data or update systems. The diagram depicts a conceptual agentic AI tool use pattern where AI system uses multiple specialized tools to process user queries by accessing various information sources. Let's look at an example implementation of the tool use pattern using Amazon Bedrock. In this architecture diagram, the tool use pattern enables fetching data from various internal and external sources and enables real-time lookup of contextual information to solve complex and realistic problems. Thanks. I'll now hand it over to Praep to continue diving into the remaining design patterns. Hello all, my name is Praep Sridan, senior solutions architect at AWS and today I'll continue from Romesh on the next two foundational AI patterns. The pattern number three is called planning pattern. Now planning pattern is an approach that allows large language models to break down complex multi-step task into smaller manageable chunks instead of attacking the problem randomly. The LM first creates a structured road map and then efficiently executes the road map adapting as it emerges. To give an example, this would be like when coding the LLM would create the overall structure of the final result instead of writing individual functions. This method ensures clarity, prevents confusion and keeps the system focused on the broader goal through execution. We can systematically think of planning as four steps. The first step is task decomposition. In task decomposition, the LLM takes a task and breaks down into manageable subtasks. And once this is done, then we do what is called a strategic structuring where systematically organizes steps for goal achievement. Then efficient execution executes a task and monitors the progress and flexible adjustments and modifies and adapts plans as needed. So by breaking large tasks into smaller task the able to provide more clarity and stay focused on the broader goal instead of getting mired into details. This is another view of the same pattern. So would first generate task and then execute to through tool use methodology then compile and provide the results back. Let's look at the multi- aent collaboration. Now this is a fairly popular and well understood pattern now but it's also probably one of the most powerful patterns available. The idea with MAC pattern or multi-agent collaboration pattern is to build the concept of delegation. You can within an organization build specialized agents. For example, as mentioned here, you can build a new modage agent and a mortgage domain knowledge base and have a supervisor agent then delegate some of the tasks to this sub agents and then get the task executed. The sub agents can operate independently but also communicate with each other to collaborate to achieve a unified goal. This pattern is especially powerful for complex use cases requiring diverse expertise, parallel processing, brainstorming such as research projects. There are several types of multi-agent patterns, right? There are collaborative agents. Each agents work on different task as part of sharing progress towards unified results. Supervised agents. A central supervisor manages and coordinates the agents verifying the outputs and supervisor tool calling. Here the supervisory agent uses LLM power tools deciding which agent to call and what argument to pass. So there are quite a few hybrid models available with multi- aent collaboration but generally speaking you know when you start constructing an agent you want to kind of think in terms of a coarse grain micros service right and look at carve out individual functionality that could be a standalone agent and then see how you can delegate it across by using a supervisor agent. If you look at some of the steps involved, we can talk about task division which is essentially the supervisor agent takes care of decomposing the complex problem into smaller task and then specialization. It identifies which agent to call for the particular task. Collaboration agents can exchange information and share results and this can happen in federally among the agents or through the supervisor and the supervisor can then combine the results to provide the final output to the user. And this is another view. You can have a finance analyzer agent a portfolio architect. And this one shows more like a collaboration pattern more than a supervised pattern because all of them are at the same level. But you get the idea having multi- aents kind of collaborating and providing a useful result to the end user. As we are building applications with generative AI, it's important to follow patterns because pattern gives you a foundational element and then how we can reuse. Thanks for listening to this video. Have a good day.