Agentic AI and its patterns

Dr. Magesh Kasthuri and Sourav Kamila, Wipro Limited

Introduction

Agentic AI, a term denoting artificial intelligence system that exhibit autonomous behavior and decision-making capabilities, is becoming increasingly prevalent across numerous industries. While the potential benefits are substantial, the adverse uses of Agentic AI in sectors such as banking, insurance, retail, manufacturing, healthcare, and learning and development highlight the need for careful consideration and regulation.

The contemporary landscape of artificial intelligence is evolving rapidly, with Agentic AI standing at the forefront of this transformation.

Earlier AI models were useful for tasks like summarizing documents and drafting emails and reports. However, the real potential is unlocked when we integrate these models with existing processes, such as databases and tools, to create systems that can solve complex problems. This approach is known as Retrieval-Augmented Generation (RAG). In RAG systems, the program follows a predefined path to answer queries, using programmatic control logic.

Alternatively, we can control the logic of a compound AI system by placing a foundation model in charge, leading to an autonomous, agentic approach. On one end of the spectrum, we instruct the foundational model or LLM to think fast and follow given instructions without deviation. On the other end, we design the system to think slowly, allowing it to create a plan, identify obstacles, adjust the plan as needed, and execute only when everything is finalized. . Please refer the below flow diagram for the same.

The multiagent pattern, involving multiple AI agents working collaboratively, offers a powerful approach to solving complex customer problems across various domains. Microsoft Azure, in conjunction with Microsoft Autogen, provides an exceptional platform to develop and deploy these multiagent solutions efficiently and effectively.

Patterns in Building Agentic AI

Four common patterns in building Agentic AI include the reflection pattern, tool use pattern, planning pattern, and multiagent pattern. Please refer the below logic flow diagram for the same.

Reflection Pattern

The reflection pattern involves AI systems that can introspect and adapt their behaviour based on feedback. It involves continuously analysing its action and outcomes to improve its performance and adapt new solutions.:

How its work-

The Reflection Pattern is a method where AI systems generate an initial response, evaluate this output for quality and correctness, and then refine the content based on their own feedback. This iterative process of generation, self-reflection, critique, and refinement leads to more accurate and polished results. By mimicking human-like feedback and revision loops, AI systems can catch mistakes, clarify ambiguities, and improve over multiple iterations. Example applications include:

· Banking: AI systems in fraud detection that continuously improve their algorithms based on detected fraud patterns.

· Healthcare: AI-driven diagnostic tools that refine their diagnostic criteria based on outcomes and feedback from medical professionals.

• Education: Intelligent tutoring systems that adapt their teaching strategies based on student performance and feedback, providing personalized learning experiences.
• Manufacturing: Predictive maintenance systems that refine their models based on equipment performance data and feedback from maintenance activities, improving reliability and reducing downtime.

Tool Use Pattern

The tool use pattern encompasses AI systems that employ various tools to achieve their objectives. It involves selecting and applying most approximate tool to achieve most specific goal. Microsoft Autogen facilitates the development of such solutions.

How its work-

The Tool Use Pattern in agentic AI enables language models to interact with external systems, transcending their limitations by accessing real-time information and specialized tools. This pattern allows AI to dynamically integrate with external resources, enhancing its capabilities beyond pre-trained data. By autonomously selecting and utilizing various tools, agentic AI can perform complex tasks such as real-time web searches, sentiment analysis, and data manipulation, demonstrating advanced problem-solving abilities. Example applications include:

· Retail: AI-powered inventory management systems use predictive analytics and IoT devices to maintain optimal stock levels. These systems analyze sales data, seasonal trends, and customer behavior to forecast demand accurately, reducing overstock and stockouts. For example, companies like Amazon and Walmart employ AI to streamline their supply chains, ensuring products are available when and where they are needed

· Manufacturing: AI systems optimize production processes through real-time data analytics and robotic automation tools. These systems monitor equipment performance, predict maintenance needs, and adjust production schedules dynamically to enhance efficiency. AI-driven predictive maintenance can reduce downtime by up to 50%, while AI-powered quality control systems ensure high standards by detecting defects in real-time

· Healthcare: AI tools assist in patient care management by analyzing medical records, predicting disease outbreaks, and personalizing treatment plans. For instance, AI-driven diagnostic tools can interpret medical images with high accuracy, aiding in early detection of conditions like cancer.

· Finance: AI systems in finance use advanced algorithms to detect fraudulent transactions, assess credit risk, and optimize investment strategies. These tools analyze vast amounts of financial data in real-time, providing insights that help institutions make informed decisions and enhance security

Planning Pattern

The planning pattern features AI that can devise complex plans to meet specific goals. It Involves defining clear objectives, breaking down task into manageable steps and then continuously evaluating progress to adapt and improve. Microsoft Autogen aids in creating planning-based agentic solutions.

How its work-

The Planning Pattern in AI involves creating detailed strategies to achieve specific objectives. This pattern starts with defining clear goals, then decomposes these goals into smaller, manageable tasks. The AI system continuously monitors progress, evaluates outcomes, and adjusts the plan as needed to ensure success. By leveraging tools like Microsoft Autogen, AI can dynamically adapt and refine its strategies, making it highly effective for complex, multi-step problem. Example applications include:

· Insurance: AI systems that plan and adjust premium calculations based on predictive models of risk and customer behaviour.

· Learning and Development: AI-driven platforms creating personalized learning plans for students based on their progress and learning styles.

• Logistics: AI systems optimize route planning and delivery schedules by analysing traffic patterns, weather conditions, and delivery constraints. This ensures timely deliveries and reduces fuel consumption. Companies like UPS and FedEx use AI to enhance their logistics operations
• Space Exploration: AI assists in mission planning for spacecraft and rovers, determining optimal paths and schedules for exploration tasks. NASA uses AI to plan and execute complex missions, ensuring efficient use of resources and maximizing scientific output
• Retail: AI-driven demand forecasting helps retailers plan inventory and promotions by analysing sales data, market trends, and consumer behaviour. This leads to better stock management and increased sales. Retail giants like Walmart and Target leverage AI for strategic planning
• Manufacturing: AI systems plan production schedules and maintenance activities by analysing equipment performance and production data. This minimizes downtime and maximizes efficiency. Companies like Siemens and General Electric use AI to optimize their manufacturing processes

Multiagent Pattern

The multiagent pattern involves multiple AI agents working collaboratively to solve problems. Microsoft Autogen supports the development of multiagent solutions.

How its work-

The Multi-Agent Pattern in AI involves multiple specialized agents working collaboratively to achieve a complex goal. Each agent is designed to handle specific tasks, and they communicate and coordinate with each other to complete the overall objective. This pattern is particularly effective for tasks that require diverse expertise and parallel processing. For example, in a software development project, different agents might handle coding, testing, and project management, working together to deliver a cohesive final product. By leveraging the strengths of individual agents, the Multi-Agent Pattern enhances efficiency, scalability, and robustness in AI system.

Example architectural flow & applications include:

· Banking: AI agents in loan processing collaborating to assess risk and approve applications efficiently.

· Healthcare: AI agents in patient care coordination, ensuring seamless communication and treatment across different healthcare providers.