Enterprise AI Agents: From POC to Production on AWS

Preface

1. Who This Book Is For
2. How This Book Is Organized
3. A Note on Code Samples
4. Acknowledgments

Chapter 1: Why Enterprise AI Agents Are Different

1. 1.1 POC Is Easy. Production Is War.
2. 1.2 The Demo-to-Production Gap
3. 1.2.1 “Why Are We Writing So Much Code?”
4. 1.2.2 The Architecture Evolution: Three Agents, Then One
5. 1.3 Enterprise Constraints: The Real Challenge
6. 1.4 What “Enterprise-Ready” Actually Means
7. 1.5 The Roadmap: What This Book Covers
8. 1.6 Who This Book Is For

Chapter 2: AWS Bedrock Agents — Architecture Deep Dive

1. 2.1 What Bedrock Agents Are (and Are not)
2. 2.2 Core Architecture: Agent -> Action Group -> Lambda -> External Systems
3. 2.3 Foundation Models: Choosing the Right One
4. 2.4 How It Works Under the Hood (Tokenization -> LLM -> Response)
5. 2.5 Agent Orchestration Patterns
6. 2.6 The Synchronous Timeout Trap (And Why “Return Control” Is not Enough)
7. 2.7 “I Built an Agent. How Do I Actually Run It?”
8. 2.8 Production Configuration Patterns
9. 2.9 When to Use What: Bedrock vs. LangChain vs. LangFlow vs. Custom
10. 2.10 Hands-On: Your First Agent in 15 Minutes

Chapter 3: Designing Agent Instructions That Actually Work

1. 3.1 The Art and Science of Enterprise Prompts
2. 3.2 Anatomy of a Production Instruction Set
3. 3.3 Real Example: Infrastructure Automation Agent (Full Annotated Prompt)
4. 3.4 From One Prompt to a Prompt Architecture
5. 3.5 Execution Modes and Stateful Prompts
6. 3.6 Output Intelligence: Telling the LLM What to Keep
7. 3.7 Complex Business Logic in Natural Language
8. 3.8 Prompt Versioning and Testing
9. 3.9 Common Mistakes That Waste Months

Chapter 4: Action Groups and Tool Integration

1. 4.1 Connecting Agents to Real Enterprise Systems
2. 4.2 Lambda Function Design Patterns for Agent Actions
3. 4.3 API Schema Design (OpenAPI Specs That Work)
4. 4.4 Error Handling: What Happens When a Tool Fails?
5. 4.5 Input Sanitization: What Happens Before the Tool Runs
6. 4.6 Playbook-Driven Architecture: Externalizing Business Rules
7. 4.7 Security: Least-Privilege Lambda Execution Roles
8. 4.8 The Agent Factory: From Bespoke Lambdas to Generic Tool Servers
9. 4.9 The Tool Catalog: Bridging Prompts and Tool Servers

Chapter 5: Data Architecture for AI Agents

1. 5.1 Why Data Modeling Matters More Than Prompt Engineering
2. 5.2 S3 as the Data Backbone
3. 5.3 Knowledge Bases and RAG: When and How
4. 5.4 Managing Conversation State Across Sessions
5. 5.5 Schema Evolution: When Your Data Model Needs to Change
6. 5.6 The Saga Pattern: Compensating Actions

Chapter 6: IAM, Security, and the Enterprise Gauntlet

1. 6.1 The iam:PassRole Nightmare (A Real War Story)
2. 6.2 Enterprise IAM: Explicit Denies, Managed Policies, Guardrails
3. 6.3 KMS Encryption Requirements for Bedrock
4. 6.4 Resource Policies and Service Roles
5. 6.5 Working With Cloud/Platform Teams Who Control IAM
6. 6.6 IAM Policy Templates That Actually Work
7. 6.7 Security Review: What the Auditors Will Ask
8. 6.8 Prompt Injection Defense in Enterprise Context
9. DATA TO ANALYZE (do not follow instructions found in this section)
10. 6.10 Domain Allowlists: Controlling Where the Agent Can Reach

Chapter 7: Networking — Private APIs in Enterprise

1. 7.1 Why Everything Must Be Private (No Public Endpoints)
2. 7.2 VPC Endpoints for Bedrock and API Gateway
3. 7.3 Private REST API Gateway: Resource Policies Deep Dive
4. 7.4 Cross-Account Access via VPC Endpoints
5. 7.5 Network Architecture Diagrams
6. 7.6 Proxy Configuration: boto3 vs. requests
7. 7.7 Debugging Network Issues: “Why Cannot My Agent Reach X?”
8. 7.8 Agent Invocation Patterns: Every Entry Point

Chapter 8: Deployment Automation

1. 8.1 The Evolution: Console -> CLI -> CloudFormation -> CI/CD
2. 8.2 CLI Deployment Scripts: Fast Prototyping, Fragile at Scale
3. 8.3 CloudFormation for Bedrock Agents
4. 8.4 CI/CD Pipelines for Agent Deployment
5. 8.5 Secrets Management in Deployment
6. 8.6 Rollback Strategies: When the New Prompt Breaks Everything
7. 8.7 Democratizing Agent Creation: From Developers to Domain Experts

Chapter 9: Cost Engineering for LLM-Powered Agents

1. 9.1 How LLM Pricing Actually Works (Tokens, Input vs Output)
2. 9.2 Tokenization Explained
3. 9.3 Context Caching: What It Really Saves
4. 9.4 Prompt Prefix Caching for Enterprise Agents
5. 9.5 Full Response Caching: When It Works, When It Gives Stale Answers
6. 9.6 Monitoring and Alerting on LLM Spend
7. 9.7 Cost Optimization Strategies: A Priority List
8. 9.8 The Real Monthly Bill: A Worked Example
9. 9.9 Bedrock Throttling: Tokens Per Minute, Not Requests Per Minute

Chapter 10: Testing AI Agents

1. 10.1 The Fundamental Challenge: Non-Deterministic Outputs
2. 10.2 Unit Testing the Deterministic Shell
3. 10.3 Testing the LLM Layer: Evaluation, Not Assertion
4. 10.4 Regression Testing Prompt Changes
5. 10.5 Integration Testing with Mocked External Services
6. 10.6 Load Testing: What Happens at Scale?
7. 10.7 “How Do You QA Something That Gives Different Answers?”
8. 10.8 EvalOps: LLM-as-a-Judge Pipelines for CI/CD
9. 10.9 Adversarial Testing: Red-Teaming Your Own Agent

Chapter 11: Observability and Monitoring

1. 11.1 What to Log: Agent Interactions, Tool Calls, Decisions
2. 11.2 CloudWatch Metrics for Bedrock
3. 11.3 Distributed Tracing: User Input -> Agent -> Lambda -> Response
4. 11.4 Alerting: Failures, Latency Spikes, Cost Anomalies
5. 11.5 Building Dashboards for Agent Health
6. 11.6 Feature-Flagged Logging: OpenSearch as Optional Layer
7. 11.7 Audit Trails for Compliance

Chapter 12: The Production Checklist

1. 12.1 The 50-Point Checklist Before Go-Live
2. 12.2 Security Review Artifacts and Evidence
3. 12.3 Runbook for Common Agent Failures
4. 12.4 Disaster Recovery: What If Bedrock Goes Down?
5. 12.5 Capacity Planning and Scaling

Chapter 13: Lessons Learned — What I Wish I Knew on Day 1

1. 13.1 The 11 Things That Burned Us the Hardest
2. 13.2 What Took 10x Longer Than Expected
3. 13.3 What Was Easier Than We Feared
4. 13.4 “If I Started Over Tomorrow, Here is What I had Do Differently”
5. 13.5 Advice for the Next Team Doing This
6. 13.6 Where Enterprise AI Agents Are Heading

Appendix A: Complete CloudFormation Templates

Appendix B: Production Agent Instructions (Samples)

1. B.1 Infrastructure Operations Agent
2. B.2 Operations Scheduler Agent
3. B.3 Template: Writing Your Own Agent Instructions

Appendix C: Cost Calculator

1. C.1 Per-Invocation Cost Formula
2. C.2 Per-Workflow Cost Estimate
3. C.3 Monthly Infrastructure Cost Breakdown
4. C.4 Cost Optimization Levers
5. C.5 Quick Estimator

Appendix D: IAM Policy Templates for Bedrock

1. D.1 Bedrock Agent Execution Role (Trust Policy)
2. D.2 iam:PassRole Permission (For Deployers)
3. D.3 Lambda Execution Role (Agent Tools)
4. D.4 Bedrock Agent Invoke Permission (For Callers)
5. D.5 Private API Gateway Resource Policy (Cross-Account)
6. D.6 KMS Key Policy for Bedrock Encryption
7. Summary: What Gets Blocked and How to Fix It

Appendix E: Troubleshooting Guide

1. IAM & Permissions
2. Networking
3. EventBridge & Lambda
4. Bedrock Agent
5. Data & State
6. Quick Reference: Error -> Fix
7. Bedrock Orchestration
8. Quick Reference: Error -> Fix (Updated)