book

Building Complex Multi-Agent Systems Using Pattern Prompting

Name: Building Complex Multi-Agent Systems Using Pattern Prompting
Author: Tim O'Brien
ISBN: 9781806114290

by Tim O'Brien

May 2026

Intermediate

310 pages

7h 26m

English

Packt Publishing

Read now

Unlock full access

Preface
Free benefits with your book
Chapter 1: Introduction: Patterns, Abstractions, and the GenAI Landscape
A vocabulary bridge: GenAI terms and their IT equivalentsStudying GenAIOn the use of abstractions and patternsHow patterns are used in this bookBenefits of breaking down GenAI microarchitectures into their constituent patternsWhat is an LLM?LLM as a continuation of big dataLLM as a badly behaving RESTful endpointHow do LLMs work?What is agentic AI?Summary
Chapter 2: Embeddings: The Language of AI
What are embeddings?What do vectors look like?What is affinity?Calculating semantic similarity programmaticallySelecting an embedding modelTesting embeddings in your codeWhat are vector databases?Chunking documentsChunking decisionsChunking strategiesSummary
Chapter 3: Building with GenAI: Parameters, Tuning, and Project Phases
Tuning GenAI systems: principles and practicesTuning the parameters of your GenAI project1) Document → Chunking2) Chunking → Retrieval3) Retrieval ↔ Prompting4) Model → Context/Tokens → Prompts5) Temperature ties to prompt behaviorConfiguring a GenAI application to answer any question about the Harry Potter booksChunking strategiesTemperatureBest practices for building a GenAI projectProject initiationIntermediate goalsCrossing the finish lineSummary
Chapter 4: Building Your First RAG App
What does "production-grade" software mean?Making technology choices: advantages of queues and topicsWhy RabbitMQ for enterprise-grade GenAI systemsQuick tour of RabbitMQBuilding a production-class RAG application using RabbitMQ with integration patternsBuilding the LLM clientTemplate MethodAdapterStrategyGenerating code and deployment configurationDeploying our solutionSummary
Chapter 5: Starting Your Data Migration Project
Identifying documents to ingest into the vector databaseSelecting the right ETL toolPlanning your data pipelineCapacity planning and throughputHandling change and continuous operationsEnsuring security, privacy, and complianceManaging parallel pipelines and CI/CDTesting and validationCost optimization and throttlingEvaluating retrieval qualityKnowledge structure and advanced retrieval (optional)Hybrid search and the limits of pure vector retrievalThe importance of data cleaning in GenAI pipelinesGraph databases and the emergence of knowledge graphsToward integrated knowledge retrievalSummary
Chapter 6: Ingesting Data Using Airbyte and Pinecone
Analysis of applicable patternsIntegration patterns implemented by ETLChannel adapterStrategy patternBuilding our ETL pipelineInstallation of Pinecone and AirbyteAnalysis of documentsOptions for configuring AirbyteConfiguring Airbyte using its no-code interfaceImporting a YAML manifestSummary
Chapter 7: Tips and Best Practices
Tips and best practices to help make your GenAI project a successBest practices for designing and operating GenAI systemsHave an R&D mindsetDon't be afraid to reason things out yourselfDiscover your own agentic patternsLead with small POCs first – keep all POCs under 2 daysSet latency and throughput targets and load test frequentlyHave a plan for managing driftDon't use data produced by GenAI with your other IT systemsValidate your UX with extensive usability testingEfficient project managementSecurity and data privacyPrompt injection attacksData leakage through context windowsHandling PII in promptsOutput validation and harmful content filteringCompliance considerations by industryBuilding a security-first cultureCost management and token budgetingBuilding a cost model earlyChoosing the right model for each taskFine-tuning as a cost optimization strategySetting cost alerts, budgets, and governanceVendor lock-in and model portabilityPrompt versioning and governanceTreat prompts as codeAdopt a versioning strategyEstablish review and approval workflowsEnable prompt rollbackDefine ownership and accountabilityEvaluation and testing frameworksBuilding your evaluation datasetAutomating evaluation pipelinesRed-teaming your applicationHandling non-determinism in testsTesting for driftSummary
Chapter 8: Pattern-Guided Coding: Using Patterns as the Design Vocabulary for GenAI Systems Built on RabbitMQ
The need for pattern-guided codingReason 1: Communication breaks down without a shared vocabularyReason 2: Design decisions become invisibleReason 3: GenAI systems fail at the integration layerHow pattern-guided coding worksBuilding your app in a four-step Topologos processPhase 1: Clarification before designPhase 2: Pattern analysis across four lensesPhase 3: Iterative approval — one decision at a timePhase 4: Final topology outputBuilding producers and consumers with GoF patternsThe Strategy pattern for producer routingThe Command pattern for agent task messagesThe Template Method pattern for consumer pipelinesThe Channel Adapter pattern for external integrationConsumer acknowledgment and the manual ACK ruleThe multi-tier dead-letter queue strategyGenerating any known GenAI pattern with a single commandGetting ready to deployThe iterative approval loopAuditing an existing systemCustomizing your GenAI applicationCase study: extending a single-LLM RAG to a dual-LLM Scatter-GatherThe before topology: single LLMThe Topologos session: proposing the changeThe new queue topology: dual LLM with Scatter-GatherWhat changed and what stayed the sameThe Aggregator's correlation windowSummary
Chapter 9: Implementing the ReAct Pattern Over RabbitMQ
Understanding the ReAct patternSetting up the RabbitMQ topologyBuilding the ReAct agentMessage schemasThe LLM call (_think)Tool dispatch (_dispatch_tool)The loop and ACK strategyImplementing tool workersError handling in tool workersScalingSending commands to the agentRunning the system end to endSetting upStart the workersSend a questionObserving the dead letter queuePreparing the system for productionSummary

Chapter 10: The Future and Limitations of LLMs
LLMs actually do: a precise descriptionThe stochastic parrot: fluency without understandingChomsky and the limits of statistical learningSearle's Chinese room: syntax is not semanticsPenrose, Gödel, and the mathematical limits of computationThe reasoning gap: what LLMs cannot reliably doSystematic mathematical reasoningMaintain logical consistency over long contextsProvide calibrated confidence in their responsesProduce novel insightsThe symbol grounding problemWhat the hype gets wrong: a taxonomy of overclaimingWhy hype is harmful: the engineering costsWhat LLMs are genuinely good atThinking about the future with clear eyesPractical guidance: working productively with known limitationsSummary
Appendix A: Pattern Reference: GoF, Enterprise Integration, Reliability, and GenAI Microarchitecture Patterns
Pattern indexPart 1 — GoF Design PatternsA.1 StrategyGoF — Behavioral | Chapters 4, 6A.2 AdapterGoF — Structural | Chapters 4, 6A.3 Template MethodGoF — Behavioral | Chapter 4Part 2 — Enterprise Integration PatternsA.4 Message ChannelEIP — Messaging Infrastructure | Chapters 4, 6A.5 Channel AdapterEIP — Messaging Infrastructure | Chapters 4, 6A.6 Publish–Subscribe ChannelEIP — Messaging | Chapter 4A.7 Content EnricherEIP — Message Transformation | Chapters 3, 4A.8 Request–ReplyEIP — Messaging | Chapter 4A.9 Correlation IdentifierEIP — Messaging | Chapter 4A.10 Scatter–GatherEIP — Message Routing | Chapter 4A.11 Dead Letter ChannelEIP — Messaging | Chapter 4A.12 Message HistoryEIP — System Management | Chapter 4A.13 Content-Based RouterEIP — Message Routing | Chapter 4A.14 Pipes and FiltersEIP — Message Routing | Chapters 5, 6Part 3 — Reliability PatternsA.15 Circuit BreakerReliability Pattern | Chapter 4A.16 Retry with Exponential BackoffReliability Pattern | Chapters 4, 7Part 4 — Microarchitecture PatternsA.17 OrchestrationMicroarchitecture Pattern | Chapters 1, 4, 7A.18 ChoreographyMicroarchitecture Pattern | Chapters 1, 7A.19 RAG MicroarchitectureGenAI Microarchitecture | Chapters 2, 3, 4, 5
Appendix B: Topologos User Manual
Quick startLoading the promptStarting a sessionConversationalPattern-firstTopology firstApprove decisions one at a timeGet your deployable artifactWhen to use TopologosCore conceptsThe two layers – pattern versus topologyCross-layer exampleThe four-phase protocolExample Phase 1 acknowledgment (fast path)Example Phase 3 decisionCommand referenceAll commands at a glanceSession commandsExample /status outputPattern layer commandsTopology layer commandsDefinition commandsOperational commandsExample /validate outputExample /simulate tool-timeout outputExample /cost outputDiagram commandsModifiersAuto-derived templatesGenerating the deployable artifactEnd-to-end workflowThe seven Phase 4 artifactsRunning /deployTargetsModifiersExamplesRendering diagrams to PNGMermaid (default)Graphviz DOTSVGPlantUMLApplying the manifest to a real brokerRoute 1 – rabbitmqadmin (CLI)Route 2 – Management HTTP APIRoute 3 – Terraform (recommended for production)Generating it inline – a typical full sessionWorked examplesExample A – Generating ReAct conversationallyPhase 2 (excerpt)Phase 3 – eight decisions (selected highlights)Phase 4 manifest excerptExample B – Composed pattern (ReAct + Tool Use)InvocationWhat changes vs.canonical ReAct?Why compose rather than extendExample C – Extending a canonical pattern (critic gate)InvocationPhase 1 acknowledgmentAdditional Phase 3 decisionsChaining multiple extensionsExample D – Defining a new pattern (Debate-and-Adjudicate)Step 1 – invoke /defineStep 2 – Topologos validates and echoes a synthetic reference cardStep 3 – compose, extend, persistWhen to define vs.extend vs.composeExample E – Defining an organization-wide topology archetypeScenarioStep 1 – define the queue archetypeStep 2 – define the DLQ topologyStep 3 – define the security baselineStep 4 – bundle them as a topology archetypeStep 5 – use it and lay a pattern on topStep 6 – persist across sessionsThe payoffRecipesTips and best practicesTroubleshootingFAQsGlossaryCanonical pattern referenceReActPlan-and-ExecuteReflectionTool UseMulti-Agent CollaborationRAGMemoryOrchestrator-SubagentHuman-in-the-LoopTree of ThoughtsTopology-layer canonical primitives
Appendix C: Unlock Your Exclusive Benefits
Unlock this Book's Free Benefits in 3 Easy Steps
Index

Content preview from Building Complex Multi-Agent Systems Using Pattern Prompting

6 Ingesting Data Using Airbyte and Pinecone

In this chapter, we will build a practical data pipeline that reads PDF documents from Amazon S3 and ingests them into the Pinecone vector database. The pipeline uses Airbyte to manage the entire extract, transform, and load (ETL) workflow.

Figure 6.1 illustrates the architecture of the pipeline we will construct. During the process, we will learn how to configure Airbyte using its low-code interface, connect it to an Amazon S3 bucket containing documents, and migrate those documents into a vector database after generating embeddings.

Figure 6.1 – Pipeline architecture: Amazon S3 → Airbyte (Load Data, ...

Become an O’Reilly member and get unlimited access to this title plus top books and audiobooks from O’Reilly and nearly 200 top publishers, thousands of courses curated by job role, 150+ live events each month,
and much more.

Read now

Unlock full access

More than 5,000 organizations count on O’Reilly

O’Reilly covers everything we've got, with content to help us build a world-class technology community, upgrade the capabilities and competencies of our teams, and improve overall team performance as well as their engagement.

Julian F.

Head of Cybersecurity

I wanted to learn C and C++, but it didn't click for me until I picked up an O'Reilly book. When I went on the O’Reilly platform, I was astonished to find all the books there, plus live events and sandboxes so you could play around with the technology.

Addison B.

Field Engineer

I’ve been on the O’Reilly platform for more than eight years. I use a couple of learning platforms, but I'm on O'Reilly more than anybody else. When you're there, you start learning. I'm never disappointed.

Amir M.

Data Platform Tech Lead

I'm always learning. So when I got on to O'Reilly, I was like a kid in a candy store. There are playlists. There are answers. There's on-demand training. It's worth its weight in gold, in terms of what it allows me to do.

Mark W.

Embedded Software Engineer

Publisher Resources

ISBN: 9781806114290

Cloud Computing

Data Engineering

Data Science

AI & ML

Programming Languages

Software Architecture

IT/Ops

Security

Design

Business

Soft Skills

Building Complex Multi-Agent Systems Using Pattern Prompting

by Tim O'Brien

6

Ingesting Data Using Airbyte and Pinecone

Become an O’Reilly member and get unlimited access to this title plus top books and audiobooks from O’Reilly and nearly 200 top publishers, thousands of courses curated by job role, 150+ live events each month,
and much more.