The field of generative Artificial Intelligence is experiencing a paradigm shift towards Agentic AI. Unlike traditional AI’s reactive, task-specific nature or generative AI’s content-generation capabilities, Agentic AI systems are characterized by their autonomy, goal-directed behavior, and capacity to learn and adapt. This represents a significant advancement in workflow automation, enabling systems to operate within defined boundaries to enhance efficiency and foster innovation.
This evolution fundamentally alters the GenAI developer’s role. Instead of prescribing precise instructions, GenAI developers become architects of autonomous decision-making spaces. Their focus shifts to defining goals, setting boundaries, and providing tools, thereby empowering the AI to determine the specific sequence of actions needed to achieve its objectives.
This workshop aims to provide practitioners with the architectural thinking and practical skills necessary to navigate this new paradigm and construct intelligent, autonomous systems.
Learning Objectives
Agentic systems become powerful by providing them with the ability to retrieve from external information, tools for interacting with their environment, and providing them with memory of the past conversations.
Upon completion, participants will be equipped to:
Architect Autonomous Systems: Move beyond simple scripts to master core architectural patterns—including ReAct, Tree of Thoughts, and Plan & Execute—to build agents capable of reasoning, reflection, and solving multistep problems.
Engineer Robust Tool-Use: Design and integrate tools that serve as the hands and eyes of an agent. Learn the principles of the Agent-Computer Interface (ACI) to ensure agents can reliably interact with APIs, databases, and other external systems.
Build with Production-Grade Frameworks: Gain hands-on experience with leading frameworks such as LangChain & LangGraph but also low-code alternatives such as Dify.ai, covering the trade-offs and practical considerations of each. We will translate theoretical knowledge into practical skills by building a complete, real-world agentic workflow.
Deploy with Confidence: Navigate the challenges of development and understandability, including techniques for rapid prototyping and stakeholder engagement.
Structure of the Workshop
Part 1: Building-blocks & Architecture of Agentic AI Systems
This section establishes the conceptual foundation for agentic systems, covering their core components and the sophisticated cognitive architectures that enable reasoning.
Fundamental Building Blocks: The workshop will detail the essential patterns for composing agentic workflows such as
Orchestration: Coordinating multiple-possibly specialized-AI agents to achieve a shared objective.
Parallelization: Executing multiple tasks concurrently to enhance system efficiency, either by dividing a large task into independent subtasks (sectioning) or by having multiple agents perform the same task for robust, aggregated results (voting).
Routing: Intelligently delegating tasks to the most appropriate specialized agent or routing the chatflow.
Evaluator-Optimizer Loops (Reflection): Establishing a self-correction loop where a generator agent iteratively refines its output based on an evaluator agent’s feedback against defined criteria.
Advanced Reasoning Architectures: The curriculum then advances to the cognitive architectures that structure an agent’s thought process. Beyond basic agent design, advanced reasoning architectures are crucial for structuring an agent’s cognitive process. These include ReAct, which connects reasoning with tool use to ground an agent’s thoughts and reduce errors. Tree of Thoughts allows agents to explore multiple reasoning paths simultaneously, mimicking human trial-and-error for complex problem-solving. For continuous improvement without fine-tuning, Reflexion enables agents to learn from past failures. Finally, Plan & Execute decouples high-level strategic planning from low-level action execution, improving performance on long, multistep tasks by maintaining a coherent strategy.
Part 2: Tool-Use & Agent-Computer Interface
This section focuses on the critical discipline of designing the interface between an agent’s reasoning engine and its ability to act in the world. Tools are what transform a reasoning model into a functional actor, allowing it to access real-time information, execute code, and interact with proprietary enterprise systems.
The central theme of this section is the ACI. Just as a Human-Computer Interface (HCI) determines the usability of traditional software for a human, the ACI determines the effectiveness of tools for an AI agent. A poorly designed ACI is a primary source of agent failure.
Part 3: Hands-on Implementation with Popular Frameworks
This section is dedicated to translating theory into practice. Participants will build a complete agentic workflow applying the introduced concepts using industry-standard frameworks, gaining practical skills that are immediately applicable in a professional context.
Navigating the Framework Ecosystem: The workshop will provide an overview of the popular frameworks, highlighting their core philosophies and ideal use cases to inform architectural decisions.
Hands-On Project: Participants will apply the concepts from Part 1 and Part 2 by building an agentic system. The project will involve defining agent roles, implementing a reasoning architecture, creating custom tools with robust ACIs and memory, and orchestrating the entire workflow. A key component of this project will be the visualization of the agent’s structure and their decision-making process, allowing participants to see how the agent operates in real-time.
Part 4: Practical Considerations for Real-World Deployment
The final part bridges the gap between a working prototype and a reliable, scalable, and trustworthy enterprise application. It addresses the pragmatic challenges that arise when deploying agentic AI in a business context.
Rapid Prototyping: The workshop will discuss tools that accelerate the development cycle, including low-code and visual interfaces. Such approaches allow for the rapid creation, management, and testing of agentic workflows through drag-and-drop interfaces, enabling faster iteration and experimentation.
Participatory Design: Furthermore, such approaches can be used to include stakeholders directly in the engineering of the agents to ensure that they solve the right problems and align with business values.
Prerequisites
The training extends on the use of Large Language Models and therefore requires basic knowledge on prompting such. However, we will do our best to provide pointers and short introductions of the key concepts involved in each part such that everybody should be able to follow. More importantly, in order to fully benefit from our practical exercises, participants should have a basic understanding of Python and feel comfortable writing smaller parts of Python code.