Agentic-Design-Patern-Ebooks on AI VOID

Dedication

Mon, 01 Jan 0001 00:00:00 +0000

To my son, Bruno,

who at two years old, brought a new and brilliant light into my life. As I explore the systems that will define our tomorrow, it is the world you will inherit that is foremost in my thoughts.

To my sons, Leonardo and Lorenzo, and my daughter Aurora,

My heart is filled with pride for the women and men you have become and the wonderful world you are building.

Acknowledgment

Mon, 01 Jan 0001 00:00:00 +0000

Acknowledgment

I would like to express my sincere gratitude to the many individuals and teams who made this book possible.

First and foremost, I thank Google for adhering to its mission, empowering Googlers, and respecting the opportunity to innovate.

I am grateful to the Office of the CTO for giving me the opportunity to explore new areas, for adhering to its mission of “practical magic,” and for its capacity to adapt to new emerging opportunities.

Foreword

Mon, 01 Jan 0001 00:00:00 +0000

Foreword

The field of artificial intelligence is at a fascinating inflection point. We are moving beyond building models that can simply process information to creating intelligent systems that can reason, plan, and act to achieve complex goals with ambiguous tasks. These “agentic” systems, as this book so aptly describes them, represent the next frontier in AI, and their development is a challenge that excites and inspires us at Google.

“Agentic Design Patterns: A Hands-On Guide to Building Intelligent Systems” arrives at the perfect moment to guide us on this journey. The book rightly points out that the power of large language models, the cognitive engines of these agents, must be harnessed with structure and thoughtful design. Just as design patterns revolutionized software engineering by providing a common language and reusable solutions to common problems, the agentic patterns in this book will be foundational for building robust, scalable, and reliable intelligent systems.

A Thought Leader's Perspective: Power and Responsibility

Mon, 01 Jan 0001 00:00:00 +0000

A Thought Leader’s Perspective: Power and Responsibility

Of all the technology cycles I’ve witnessed over the past four decades—from the birth of the personal computer and the web, to the revolutions in mobile and cloud—none has felt quite like this one. For years, the discourse around Artificial Intelligence was a familiar rhythm of hype and disillusionment, the so-called “AI summers” followed by long, cold winters. But this time, something is different. The conversation has palpably shifted. If the last eighteen months were
about the engine—the breathtaking, almost vertical ascent of Large Language Models (LLMs)—the next era will be about the car we build around it. It will be about the frameworks that harness this raw power, transforming it from a generator of plausible text into a true agent of action.

Introduction

Mon, 01 Jan 0001 00:00:00 +0000

Preface

Welcome to “Agentic Design Patterns: A Hands-On Guide to Building Intelligent Systems.” As we look across the landscape of modern artificial intelligence, we see a clear evolution from simple, reactive programs to sophisticated, autonomous entities capable of understanding context, making decisions, and interacting dynamically with their environment and other systems. These are the intelligent agents and the agentic systems they comprise.

The advent of powerful large language models (LLMs) has provided unprecedented capabilities for understanding and generating human-like content such as text and media, serving as the cognitive engine for many of these agents. However, orchestrating these capabilities into systems that can reliably achieve complex goals requires more than just a powerful model. It requires structure, design, and a thoughtful approach to how the agent perceives, plans, acts, and interacts.

What makes an AI system an "agent"?

Mon, 01 Jan 0001 00:00:00 +0000

What makes an AI system an Agent?

In simple terms, an AI agent is a system designed to perceive its environment and take actions to achieve a specific goal. It’s an evolution from a standard Large Language Model (LLM), enhanced with the abilities to plan, use tools, and interact with its surroundings. Think of an Agentic AI as a smart assistant that learns on the job. It follows a simple, five-step loop to get things done (see Fig.1):

Chapter 1: Prompt Chaining

Mon, 01 Jan 0001 00:00:00 +0000

Chapter 1: Prompt Chaining

Prompt Chaining Pattern Overview

Prompt chaining, sometimes referred to as Pipeline pattern, represents a powerful paradigm for handling intricate tasks when leveraging large language models (LLMs). Rather than expecting an LLM to solve a complex problem in a single, monolithic step, prompt chaining advocates for a divide-and-conquer strategy. The core idea is to break down the original, daunting problem into a sequence of smaller, more manageable sub-problems. Each sub-problem is addressed individually through a specifically designed prompt, and the output generated from one prompt is strategically fed as input into the subsequent prompt in the chain.

Chapter 2: Routing

Mon, 01 Jan 0001 00:00:00 +0000

Chapter 2: Routing

Routing Pattern Overview

While sequential processing via prompt chaining is a foundational technique for executing deterministic, linear workflows with language models, its applicability is limited in scenarios requiring adaptive responses. Real-world agentic systems must often arbitrate between multiple potential actions based on contingent factors, such as the state of the environment, user input, or the outcome of a preceding operation. This capacity for dynamic decision-making, which governs the flow of control to different specialized functions, tools, or sub-processes, is achieved through a mechanism known as routing.

Chapter 3: Parallelization

Mon, 01 Jan 0001 00:00:00 +0000

Chapter 3: Parallelization

Parallelization Pattern Overview

In the previous chapters, we’ve explored Prompt Chaining for sequential workflows and Routing for dynamic decision-making and transitions between different paths. While these patterns are essential, many complex agentic tasks involve multiple sub-tasks that can be executed simultaneously rather than one after another. This is where the Parallelization pattern becomes crucial.

Parallelization involves executing multiple components, such as LLM calls, tool usages, or even entire sub-agents, concurrently (see Fig.1). Instead of waiting for one step to complete before starting the next, parallel execution allows independent tasks to run at the same time, significantly reducing the overall execution time for tasks that can be broken down into independent parts.

Chapter 4: Reflection

Mon, 01 Jan 0001 00:00:00 +0000

Chapter 4: Reflection

Reflection Pattern Overview

In the preceding chapters, we’ve explored fundamental agentic patterns: Chaining for sequential execution, Routing for dynamic path selection, and Parallelization for concurrent task execution. These patterns enable agents to perform complex tasks more efficiently and flexibly. However, even with sophisticated workflows, an agent’s initial output or plan might not be optimal, accurate, or complete. This is where the Reflection pattern comes into play.

The Reflection pattern involves an agent evaluating its own work, output, or internal state and using that evaluation to improve its performance or refine its response. It’s a form of self-correction or self-improvement, allowing the agent to iteratively refine its output or adjust its approach based on feedback, internal critique, or comparison against desired criteria. Reflection can occasionally be facilitated by a separate agent whose specific role is to analyze the output of an initial agent.

Chapter 5: Tool Use

Mon, 01 Jan 0001 00:00:00 +0000

Chapter 5: Tool Use (Function Calling)

Tool Use Pattern Overview

So far, we’ve discussed agentic patterns that primarily involve orchestrating interactions between language models and managing the flow of information within the agent’s internal workflow (Chaining, Routing, Parallelization, Reflection). However, for agents to be truly useful and interact with the real world or external systems, they need the ability to use Tools.

The Tool Use pattern, often implemented through a mechanism called Function Calling, enables an agent to interact with external APIs, databases, services, or even execute code. It allows the LLM at the core of the agent to decide when and how to use a specific external function based on the user’s request or the current state of the task.

Chapter 6: Planning

Mon, 01 Jan 0001 00:00:00 +0000

Chapter 6: Planning

Intelligent behavior often involves more than just reacting to the immediate input. It requires foresight, breaking down complex tasks into smaller, manageable steps, and strategizing how to achieve a desired outcome. This is where the Planning pattern comes into play. At its core, planning is the ability for an agent or a system of agents to formulate a sequence of actions to move from an initial state towards a goal state.

Chapter 7: Multi-Agent

Mon, 01 Jan 0001 00:00:00 +0000

Chapter 7: Multi-Agent Collaboration

While a monolithic agent architecture can be effective for well-defined problems, its capabilities are often constrained when faced with complex, multi-domain tasks. The Multi-Agent Collaboration pattern addresses these limitations by structuring a system as a cooperative ensemble of distinct, specialized agents. This approach is predicated on the principle of task decomposition, where a high-level objective is broken down into discrete sub-problems. Each sub-problem is then assigned to an agent possessing the specific tools, data access, or reasoning capabilities best suited for that task.

Chapter 8: Memory Management

Mon, 01 Jan 0001 00:00:00 +0000

Chapter 8: Memory Management

Effective memory management is crucial for intelligent agents to retain information. Agents require different types of memory, much like humans, to operate efficiently. This chapter delves into memory management, specifically addressing the immediate (short-term) and persistent (long-term) memory requirements of agents.

In agent systems, memory refers to an agent’s ability to retain and utilize information from past interactions, observations, and learning experiences. This capability allows agents to make informed decisions, maintain conversational context, and improve over time. Agent memory is generally categorized into two main types:

Chapter 9: Learning and Adaptation

Mon, 01 Jan 0001 00:00:00 +0000

Chapter 9: Learning and Adaptation

Learning and adaptation are pivotal for enhancing the capabilities of artificial intelligence agents. These processes enable agents to evolve beyond predefined parameters, allowing them to improve autonomously through experience and environmental interaction. By learning and adapting, agents can effectively manage novel situations and optimize their performance without constant manual intervention. This chapter explores the principles and mechanisms underpinning agent learning and adaptation in detail.

The big picture

Agents learn and adapt by changing their thinking, actions, or knowledge based on new experiences and data. This allows agents to evolve from simply following instructions to becoming smarter over time.

Chapter 10: Model Context Protocol (MCP)

Mon, 01 Jan 0001 00:00:00 +0000

Chapter 10: Model Context Protocol

To enable LLMs to function effectively as agents, their capabilities must extend beyond multimodal generation. Interaction with the external environment is necessary, including access to current data, utilization of external software, and execution of specific operational tasks. The Model Context Protocol (MCP) addresses this need by providing a standardized interface for LLMs to interface with external resources. This protocol serves as a key mechanism to facilitate consistent and predictable integration.

Chapter 11: Goal Setting and Monitoring

Mon, 01 Jan 0001 00:00:00 +0000

Chapter 11: Goal Setting and Monitoring

For AI agents to be truly effective and purposeful, they need more than just the ability to process information or use tools; they need a clear sense of direction and a way to know if they’re actually succeeding. This is where the Goal Setting and Monitoring pattern comes into play. It’s about giving agents specific objectives to work towards and equipping them with the means to track their progress and determine if those objectives have been met.

Chapter 12: Exception Handling and Recovery

Mon, 01 Jan 0001 00:00:00 +0000

Chapter 12: Exception Handling and Recovery

For AI agents to operate reliably in diverse real-world environments, they must be able to manage unforeseen situations, errors, and malfunctions. Just as humans adapt to unexpected obstacles, intelligent agents need robust systems to detect problems, initiate recovery procedures, or at least ensure controlled failure. This essential requirement forms the basis of the Exception Handling and Recovery pattern.

This pattern focuses on developing exceptionally durable and resilient agents that can maintain uninterrupted functionality and operational integrity despite various difficulties and anomalies. It emphasizes the importance of both proactive preparation and reactive strategies to ensure continuous operation, even when facing challenges. This adaptability is critical for agents to function successfully in complex and unpredictable settings, ultimately boosting their overall effectiveness and trustworthiness.

Chapter 13: Human-in-the-Loop

Mon, 01 Jan 0001 00:00:00 +0000

Chapter 13: Human-in-the-Loop

The Human-in-the-Loop (HITL) pattern represents a pivotal strategy in the development and deployment of Agents. It deliberately interweaves the unique strengths of human cognition—such as judgment, creativity, and nuanced understanding—with the computational power and efficiency of AI. This strategic integration is not merely an option but often a necessity, especially as AI systems become increasingly embedded in critical decision-making processes.

The core principle of HITL is to ensure that AI operates within ethical boundaries, adheres to safety protocols, and achieves its objectives with optimal effectiveness. These concerns are particularly acute in domains characterized by complexity, ambiguity, or significant risk, where the implications of AI errors or misinterpretations can be substantial. In such scenarios, full autonomy—where AI systems function independently without any human intervention—may prove to be imprudent. HITL acknowledges this reality and emphasizes that even with rapidly advancing AI technologies, human oversight, strategic input, and collaborative interactions remain indispensable.

Chapter 14: Knowledge Retrieval (RAG)

Mon, 01 Jan 0001 00:00:00 +0000

Chapter 14: Knowledge Retrieval (RAG)

LLMs exhibit substantial capabilities in generating human-like text. However, their knowledge base is typically confined to the data on which they were trained, limiting their access to real-time information, specific company data, or highly specialized details. Knowledge Retrieval (RAG, or Retrieval Augmented Generation), addresses this limitation. RAG enables LLMs to access and integrate external, current, and context-specific information, thereby enhancing the accuracy, relevance, and factual basis of their outputs.

Chapter 15: Inter-Agent Communication (A2A)

Mon, 01 Jan 0001 00:00:00 +0000

Chapter 15: Inter-Agent Communication (A2A)

Individual AI agents often face limitations when tackling complex, multifaceted problems, even with advanced capabilities. To overcome this, Inter-Agent Communication (A2A) enables diverse AI agents, potentially built with different frameworks, to collaborate effectively. This collaboration involves seamless coordination, task delegation, and information exchange.

Google’s A2A protocol is an open standard designed to facilitate this universal communication. This chapter will explore A2A, its practical applications, and its implementation within the Google ADK.

Chapter 16: Resource-Aware Optimization

Mon, 01 Jan 0001 00:00:00 +0000

Chapter 16: Resource-Aware Optimization

Resource-Aware Optimization enables intelligent agents to dynamically monitor and manage computational, temporal, and financial resources during operation. This differs from simple planning, which primarily focuses on action sequencing. Resource-Aware Optimization requires agents to make decisions regarding action execution to achieve goals within specified resource budgets or to optimize efficiency. This involves choosing between more accurate but expensive models and faster, lower-cost ones, or deciding whether to allocate additional compute for a more refined response versus returning a quicker, less detailed answer.

Chapter 17: Reasoning Techniques

Mon, 01 Jan 0001 00:00:00 +0000

Chapter 17: Reasoning Techniques

This chapter delves into advanced reasoning methodologies for intelligent agents, focusing on multi-step logical inferences and problem-solving. These techniques go beyond simple sequential operations, making the agent’s internal reasoning explicit. This allows agents to break down problems, consider intermediate steps, and reach more robust and accurate conclusions. A core principle among these advanced methods is the allocation of increased computational resources during inference. This means granting the agent, or the underlying LLM, more processing time or steps to process a query and generate a response. Rather than a quick, single pass, the agent can engage in iterative refinement, explore multiple solution paths, or utilize external tools. This extended processing time during inference often significantly enhances accuracy, coherence, and robustness, especially for complex problems requiring deeper analysis and deliberation.

Chapter 18: Guardrails/Safety Patterns

Mon, 01 Jan 0001 00:00:00 +0000

Chapter 18: Guardrails/Safety Patterns

Guardrails, also referred to as safety patterns, are crucial mechanisms that ensure intelligent agents operate safely, ethically, and as intended, particularly as these agents become more autonomous and integrated into critical systems. They serve as a protective layer, guiding the agent’s behavior and output to prevent harmful, biased, irrelevant, or otherwise undesirable responses. These guardrails can be implemented at various stages, including Input Validation/Sanitization to filter malicious content, Output Filtering/Post-processing to analyze generated responses for toxicity or bias, Behavioral Constraints (Prompt-level) through direct instructions, Tool Use Restrictions to limit agent capabilities, External Moderation APIs for content moderation, and Human Oversight/Intervention via “Human-in-the-Loop” mechanisms.

Chapter 19: Evaluation and Monitoring

Mon, 01 Jan 0001 00:00:00 +0000

Chapter 19: Evaluation and Monitoring

This chapter examines methodologies that allow intelligent agents to systematically assess their performance, monitor progress toward goals, and detect operational anomalies. While Chapter 11 outlines goal setting and monitoring, and Chapter 17 addresses Reasoning mechanisms, this chapter focuses on the continuous, often external, measurement of an agent’s effectiveness, efficiency, and compliance with requirements. This includes defining metrics, establishing feedback loops, and implementing reporting systems to ensure agent performance aligns with expectations in operational environments (see Fig.1)

Chapter 20: Prioritization

Mon, 01 Jan 0001 00:00:00 +0000

Chapter 20: Prioritization

In complex, dynamic environments, Agents frequently encounter numerous potential actions, conflicting goals, and limited resources. Without a defined process for determining the subsequent action, the agents may experience reduced efficiency, operational delays, or failures to achieve key objectives. The prioritization pattern addresses this issue by enabling agents to assess and rank tasks, objectives, or actions based on their significance, urgency, dependencies, and established criteria. This ensures the agents concentrate efforts on the most critical tasks, resulting in enhanced effectiveness and goal alignment.

Chapter 21: Exploration and Discovery

Mon, 01 Jan 0001 00:00:00 +0000

Chapter 21: Exploration and Discovery

This chapter explores patterns that enable intelligent agents to actively seek out novel information, uncover new possibilities, and identify unknown unknowns within their operational environment. Exploration and discovery differ from reactive behaviors or optimization within a predefined solution space. Instead, they focus on agents proactively venturing into unfamiliar territories, experimenting with new approaches, and generating new knowledge or understanding. This pattern is crucial for agents operating in open-ended, complex, or rapidly evolving domains where static knowledge or pre-programmed solutions are insufficient. It emphasizes the agent’s capacity to expand its understanding and capabilities.

Appendix A: Advanced Prompting Techniques

Mon, 01 Jan 0001 00:00:00 +0000

Appendix A: Advanced Prompting Techniques

Introduction to Prompting

Prompting, the primary interface for interacting with language models, is the process of crafting inputs to guide the model towards generating a desired output. This involves structuring requests, providing relevant context, specifying the output format, and demonstrating expected response types. Well-designed prompts can maximize the potential of language models, resulting in accurate, relevant, and creative responses. In contrast, poorly designed prompts can lead to ambiguous, irrelevant, or erroneous outputs.

Appendix B - AI Agentic ….: From GUI to Real world environment

Mon, 01 Jan 0001 00:00:00 +0000

Appendix B - AI Agentic Interactions: From GUI to Real World environment

AI agents are increasingly performing complex tasks by interacting with digital interfaces and the physical world. Their ability to perceive, process, and act within these varied environments is fundamentally transforming automation, human-computer interaction, and intelligent systems. This appendix explores how agents interact with computers and their environments, highlighting advancements and projects.

Interaction: Agents with Computers

The evolution of AI from conversational partners to active, task-oriented agents is being driven by Agent-Computer Interfaces (ACIs). These interfaces allow AI to interact directly with a computer’s Graphical User Interface (GUI), enabling it to perceive and manipulate visual elements like icons and buttons just as a human would. This new method moves beyond the rigid, developer-dependent scripts of traditional automation that relied on APIs and system calls. By using the visual “front door” of software, AI can now automate complex digital tasks in a more flexible and powerful way, a process that involves several key stages:

Appendix C - Quick overview of Agentic Frameworks

Mon, 01 Jan 0001 00:00:00 +0000

Appendix C - Quick overview of Agentic Frameworks

LangChain

LangChain is a framework for developing applications powered by LLMs. Its core strength lies in its LangChain Expression Language (LCEL), which allows you to “pipe” components together into a chain. This creates a clear, linear sequence where the output of one step becomes the input for the next. It’s built for workflows that are Directed Acyclic Graphs (DAGs), meaning the process flows in one direction without loops.

Appendix D - Building an Agent with AgentSpace (on-line only)

Mon, 01 Jan 0001 00:00:00 +0000

Appendix D - Building an Agent with AgentSpace

Overview

AgentSpace is a platform designed to facilitate an “agent-driven enterprise” by integrating artificial intelligence into daily workflows. At its core, it provides a unified search capability across an organization’s entire digital footprint, including documents, emails, and databases. This system utilizes advanced AI models, like Google’s Gemini, to comprehend and synthesize information from these varied sources.

The platform enables the creation and deployment of specialized AI “agents” that can perform complex tasks and automate processes. These agents are not merely chatbots; they can reason, plan, and execute multi-step actions autonomously. For instance, an agent could research a topic, compile a report with citations, and even generate an audio summary.

Appendix E - AI Agents on the CLI (online)

Mon, 01 Jan 0001 00:00:00 +0000

Appendix E - AI Agents on the CLI

Introduction

The developer’s command line, long a bastion of precise, imperative commands, is undergoing a profound transformation. It is evolving from a simple shell into an intelligent, collaborative workspace powered by a new class of tools: AI Agent Command-Line Interfaces (CLIs). These agents move beyond merely executing commands; they understand natural language, maintain context about your entire codebase, and can perform complex, multi-step tasks that automate significant parts of the development lifecycle.

Appendix F - Under the Hood: An Inside Look at the Agents’ Reasoning Engines

Mon, 01 Jan 0001 00:00:00 +0000

Appendix F - Under the Hood: An Inside Look at the Agents’Reasoning Engines

The emergence of intelligent Agents represents a pivotal shift in artificial intelligence. These are systems designed to plan, strategize, and execute complex tasks, and at the cognitive core of each lies a LLM. This LLM is not merely a sophisticated text generator; it serves as the foundational reasoning engine, the central “mind” that empowers the Agent to make decisions, formulate plans, and interact with its environment.

Appendix G - Coding agents

Mon, 01 Jan 0001 00:00:00 +0000

Appendix G - Coding Agents

Vibe Coding: A Starting Point

“Vibe coding” has become a powerful technique for rapid innovation and creative exploration. This practice involves using LLMs to generate initial drafts, outline complex logic, or build quick prototypes, significantly reducing initial friction. It is invaluable for overcoming the “blank page” problem, enabling developers to quickly transition from a vague concept to tangible, runnable code. Vibe coding is particularly effective when exploring unfamiliar APIs or testing novel architectural patterns, as it bypasses the immediate need for perfect implementation. The generated code often acts as a creative catalyst, providing a foundation for developers to critique, refactor, and expand upon. Its primary strength lies in its ability to accelerate the initial discovery and ideation phases of the software lifecycle. However, while vibe coding excels at brainstorming, developing robust, scalable, and maintainable software demands a more structured approach, shifting from pure generation to a collaborative partnership with specialized coding agents.

Conclusion

Mon, 01 Jan 0001 00:00:00 +0000

Conclusion

Throughout this book we have journeyed from the foundational concepts of agentic AI to the practical implementation of sophisticated, autonomous systems. We began with the premise that building intelligent agents is akin to creating a complex work of art on a technical canvas—a process that requires not just a powerful cognitive engine like a large language model, but also a robust set of architectural blueprints. These blueprints, or agentic patterns, provide the structure and reliability needed to transform simple, reactive models into proactive, goal-oriented entities capable of complex reasoning and action.

Glossary

Mon, 01 Jan 0001 00:00:00 +0000

Glossary

Fundamental Concepts

Prompt: A prompt is the input, typically in the form of a question, instruction, or statement, that a user provides to an AI model to elicit a response. The quality and structure of the prompt heavily influence the model’s output, making prompt engineering a key skill for effectively using AI.

Context Window: The context window is the maximum number of tokens an AI model can process at once, including both the input and its generated output. This fixed size is a critical limitation, as information outside the window is ignored, while larger windows enable more complex conversations and document analysis.

Index of Terms

Mon, 01 Jan 0001 00:00:00 +0000

Glossary

Fundamental Concepts

Prompt: A prompt is the input, typically in the form of a question, instruction, or statement, that a user provides to an AI model to elicit a response. The quality and structure of the prompt heavily influence the model’s output, making prompt engineering a key skill for effectively using AI.

Context Window: The context window is the maximum number of tokens an AI model can process at once, including both the input and its generated output. This fixed size is a critical limitation, as information outside the window is ignored, while larger windows enable more complex conversations and document analysis.

In-Context Learning: In-context learning is an AI’s ability to learn a new task from examples provided directly in the prompt, without requiring any retraining. This powerful feature allows a single, general-purpose model to be adapted to countless specific tasks on the fly.

Zero-Shot, One-Shot, & Few-Shot Prompting: These are prompting techniques where a model is given zero, one, or a few examples of a task to guide its response. Providing more examples generally helps the model better understand the user’s intent and improves its accuracy for the specific task.

Multimodality: Multimodality is an AI’s ability to understand and process information across multiple data types like text, images, and audio. This allows for more versatile and human-like interactions, such as describing an image or answering a spoken question.

Grounding: Grounding is the process of connecting a model’s outputs to verifiable, real-world information sources to ensure factual accuracy and reduce hallucinations. This is often achieved with techniques like RAG to make AI systems more trustworthy.

Core AI Model Architectures

Transformers: The Transformer is the foundational neural network architecture for most modern LLMs. Its key innovation is the self-attention mechanism, which efficiently processes long sequences of text and captures complex relationships between words.

Recurrent Neural Network (RNN): The Recurrent Neural Network is a foundational architecture that preceded the Transformer. RNNs process information sequentially, using loops to maintain a “memory” of previous inputs, which made them suitable for tasks like text and speech processing.

Mixture of Experts (MoE): Mixture of Experts is an efficient model architecture where a “router” network dynamically selects a small subset of “expert” networks to handle any given input. This allows models to have a massive number of parameters while keeping computational costs manageable.

Diffusion Models: Diffusion models are generative models that excel at creating high-quality images. They work by adding random noise to data and then training a model to meticulously reverse the process, allowing them to generate novel data from a random starting point.

Mamba: Mamba is a recent AI architecture using a Selective State Space Model (SSM) to process sequences with high efficiency, especially for very long contexts. Its selective mechanism allows it to focus on relevant information while filtering out noise, making it a potential alternative to the Transformer.

The LLM Development Lifecycle

The development of a powerful language model follows a distinct sequence. It begins with Pre-training, where a massive base model is built by training it on a vast dataset of general internet text to learn language, reasoning, and world knowledge. Next is Fine-tuning, a specialization phase where the general model is further trained on smaller, task-specific datasets to adapt its capabilities for a particular purpose. The final stage is Alignment, where the specialized model’s behavior is adjusted to ensure its outputs are helpful, harmless, and aligned with human values.

Pre-training Techniques: Pre-training is the initial phase where a model learns general knowledge from vast amounts of data. The top techniques for this involve different objectives for the model to learn from. The most common is Causal Language Modeling (CLM), where the model predicts the next word in a sentence. Another is Masked Language Modeling (MLM), where the model fills in intentionally hidden words in a text. Other important methods include Denoising Objectives, where the model learns to restore a corrupted input to its original state, Contrastive Learning, where it learns to distinguish between similar and dissimilar pieces of data, and Next Sentence Prediction (NSP), where it determines if two sentences logically follow each other.

Fine-tuning Techniques: Fine-tuning is the process of adapting a general pre-trained model to a specific task using a smaller, specialized dataset. The most common approach is Supervised Fine-Tuning (SFT), where the model is trained on labeled examples of correct input-output pairs. A popular variant is Instruction Tuning, which focuses on training the model to better follow user commands. To make this process more efficient, Parameter-Efficient Fine-Tuning (PEFT) methods are used, with top techniques including LoRA (Low-Rank Adaptation), which only updates a small number of parameters, and its memory-optimized version, QLoRA. Another technique, Retrieval-Augmented Generation (RAG), enhances the model by connecting it to an external knowledge source during the fine-tuning or inference stage.

Alignment & Safety Techniques: Alignment is the process of ensuring an AI model’s behavior aligns with human values and expectations, making it helpful and harmless. The most prominent technique is Reinforcement Learning from Human Feedback (RLHF), where a “reward model” trained on human preferences guides the AI’s learning process, often using an algorithm like Proximal Policy Optimization (PPO) for stability. Simpler alternatives have emerged, such as Direct Preference Optimization (DPO), which bypasses the need for a separate reward model, and Kahneman-Tversky Optimization (KTO), which simplifies data collection further. To ensure safe deployment, Guardrails are implemented as a final safety layer to filter outputs and block harmful actions in real-time.

Enhancing AI Agent Capabilities

AI agents are systems that can perceive their environment and take autonomous actions to achieve goals. Their effectiveness is enhanced by robust reasoning frameworks.

Chain of Thought (CoT): This prompting technique encourages a model to explain its reasoning step-by-step before giving a final answer. This process of “thinking out loud” often leads to more accurate results on complex reasoning tasks.

Tree of Thoughts (ToT): Tree of Thoughts is an advanced reasoning framework where an agent explores multiple reasoning paths simultaneously, like branches on a tree. It allows the agent to self-evaluate different lines of thought and choose the most promising one to pursue, making it more effective at complex problem-solving.

ReAct (Reason and Act): ReAct is an agent framework that combines reasoning and acting in a loop. The agent first “thinks” about what to do, then takes an “action” using a tool, and uses the resulting observation to inform its next thought, making it highly effective at solving complex tasks.

Planning: This is an agent’s ability to break down a high-level goal into a sequence of smaller, manageable sub-tasks. The agent then creates a plan to execute these steps in order, allowing it to handle complex, multi-step assignments.

Deep Research: Deep research refers to an agent’s capability to autonomously explore a topic in-depth by iteratively searching for information, synthesizing findings, and identifying new questions. This allows the agent to build a comprehensive understanding of a subject far beyond a single search query.

Critique Model: A critique model is a specialized AI model trained to review, evaluate, and provide feedback on the output of another AI model. It acts as an automated critic, helping to identify errors, improve reasoning, and ensure the final output meets a desired quality standard.

Index of Terms

This index of terms was generated using Gemini Pro 2.5. The prompt and reasoning steps are included at the end to demonstrate the time-saving benefits and for educational purposes.

Online Contribution - Frequently Asked Questions: Agentic Design Patterns

Mon, 01 Jan 0001 00:00:00 +0000

Frequently Asked Questions: Agentic Design Patterns

What is an “agentic design pattern”? An agentic design pattern is a reusable, high-level solution to a common problem encountered when building intelligent, autonomous systems (agents). These patterns provide a structured framework for designing agent behaviors, much like software design patterns do for traditional programming. They help developers build more robust, predictable, and effective AI agents.

What is the main goal of this guide? The guide aims to provide a practical, hands-on introduction to designing and building agentic systems. It moves beyond theoretical discussions to offer concrete architectural blueprints that developers can use to create agents capable of complex, goal-oriented behavior in a reliable way.