The New Rules of Building Reliable AI Agents

Introduction

Chapter 1: What We Mean by AI Agents

1. From Chatbots to Agents: A Taxonomy
2. The Reliability Gap: Why Agent Behavior Is Fundamentally Different
3. Deterministic Code Meets Probabilistic Reasoning
4. The Core Ingredients of an Agent System
5. Common Misconceptions That Cause Real Outages

Chapter 2: Why Agents Fail

1. Hallucinations and False Confidence
2. Reasoning Errors and Chain-of-Thought Fragility
3. Non-Determinism and Unreproducible Behavior
4. Tool Misuse: Wrong Tools, Wrong Parameters, Wrong Time
5. Infinite Loops, Divergence and Resource Exhaustion
6. Cascading Failures and State Corruption
7. Security Failures: Prompt Injection, Leakage and Abuse
8. Partial Execution and Incomplete Work

Chapter 3: The Reliability Stack: A Layered Model

1. The Layers: Control, Reasoning, Tools, State, Observability
2. Where Determinism Lives (and Where It Must Not)
3. Explicit Contracts and Invariants
4. Bounded Autonomy as a Design Principle
5. Failure Containment: Blame Lines and Isolation

Chapter 4: Control Architectures: Patterns and Trade-offs

1. The Simple Agent Loop: When It Works, When It Breaks
2. Deterministic State Machines Around Probabilistic Transitions
3. Graph-Based Workflows and DAG Orchestration
4. Hierarchical Agents and Subagent Delegation
5. The Supervisor Pattern: Meta-Agent Coordination
6. Decision Criteria: Choosing an Architecture for Your Constraints

Chapter 5: Planning, Decomposition and Task Management

1. Task Decomposition Strategies: Recursive, Iterative and Hybrid
2. Plan vs. Execution: Separation of Concerns
3. Dynamic Replanning and Adaptation
4. Goal Specification: From Vague Intentions to Machine-Checkable Conditions
5. Monitoring Plan Health and Detecting Derailment

Chapter 6: Tool Design and Integration

1. The Tool Contract: Typed Inputs, Validated Outputs, Explicit Errors
2. Idempotency and Safe Retries
3. Tool Discovery and Selection Reliability
4. Parameter Validation and Sanitization
5. Asynchronous Tools and Long-Running Operations
6. Tool Observability: Logging, Timing and Metrics

Chapter 7: Context Engineering and Memory Systems

1. The Context Window as a Constrained Resource
2. Short-Term Memory: Conversation History and Working Context
3. Long-Term Memory: Storage, Indexing and Retrieval
4. RAG Reliability: Freshness, Accuracy and Citation
5. Context Compression and Summarization Risks
6. Memory Consistency and Concurrency

Chapter 8: Structured Outputs and Grounding

1. Schemas, Grammars and Constrained Decoding
2. Output Validation: Pre-flight and Post-flight Checks
3. Grounding Outputs in Verifiable Evidence
4. Confidence, Uncertainty and When to Fail Open
5. Multi-Step Verification and Cross-Check Patterns

Chapter 9: Multi-Agent Systems and Coordination

1. When You Actually Need Multiple Agents
2. Coordination Patterns: Messaging, Shared State, Hierarchical
3. Communication Reliability and Protocol Design
4. Conflict Resolution and Deadlock
5. Emergent Behavior and Unintended Interactions
6. Multi-Agent Security and Isolation

Chapter 10: Building a Reliable Agent: Reference Architecture I (Single Agent)

1. System Overview and Design Goals
2. Project Structure and Dependencies
3. Typed Data Models and Configuration
4. Model Integration and Prompt Management
5. Tool Definitions with Validation and Error Handling
6. State Persistence and Recovery
7. The Control Loop: Control Flow Around Probabilistic Reasoning
8. Automated Tests and Basic Evaluation

Chapter 11: Reference Architecture II (Workflow-Based Agent)

1. When Workflows Beat Pure Autonomy
2. Defining the Workflow Graph
3. Step Implementations: Mixing Determinism and AI
4. Error Handling and Recovery in Workflows
5. Human Approval Gates
6. Observability Across Workflow Executions
7. Testing Workflow Behavior

Chapter 12: Reference Architecture III (Multi-Agent System)

1. Architecture Overview: Roles, Responsibilities, Boundaries
2. Message Protocol and Communication Layer
3. Agent Implementations with Encapsulated Reasoning
4. Coordination and Orchestration Logic
5. Shared State and Consistency
6. Security: Isolation, Authorization, Least Privilege
7. End-to-End Observability
8. Deployment, Scaling and Operations
9. Complete Evaluation and Regression Suite

Chapter 13: Evaluation: Testing Agents Like Real Systems

1. The Test Pyramid for Agentic Systems
2. Unit Testing: Isolated Components and Tools
3. Integration Testing: Agent in a Controlled Environment
4. Scenario-Based Testing and Simulations
5. Adversarial Testing and Red Teaming
6. Deterministic Simulations and Mock Environments
7. Regression Testing for Model and Prompt Changes

Chapter 14: Model-Graded Evaluation and Its Limits

1. How Model Grading Works
2. Designing Effective Grading Prompts
3. Calibration and Agreement with Human Judgment
4. Limitations: Bias, Fragility and Manipulation
5. Combining Model Grading with Deterministic Checks
6. When Human Evaluation Is Non-Negotiable

Chapter 15: Observability, Monitoring and Incident Response

1. What to Observe: Traces, Logs, Metrics and Audits
2. Distributed Tracing for Agent Executions
3. Structured Logging: Machine-Readable Narratives
4. Key Metrics for Agent Reliability
5. Alerting on Agent Behavior: Signals vs. Noise
6. Incident Response: Diagnosing Agent Failures
7. Postmortems and Continuous Improvement

Chapter 16: Security Engineering for AI Agents

1. The Expanded Attack Surface of Agents
2. Prompt Injection: Types, Defenses and Residual Risks
3. Tool Authorization and Least Privilege
4. Data Leakage and Privacy Preservation
5. Sandboxing and Execution Isolation
6. Input/Output Filtering and Sanitization
7. Security Testing for Agent Systems

Chapter 17: Production Deployment and Infrastructure

1. Deployment Architectures: Monolith, Microservices, Serverless
2. Model Access: Direct API, Proxy and Routing
3. Concurrency and Rate Limiting
4. Caching Strategies: Prompts, Responses and Tool Results
5. Token Economics and Cost Attribution
6. Capacity Planning and Load Management
7. Graceful Degradation and Fallback

Chapter 18: Governance, Compliance and the Future

1. Change Management: Models, Prompts and Tools
2. Compliance and Regulatory Considerations
3. Documentation and Runbooks for Agent Systems
4. Reproducibility and Experimentation Tracking
5. Open Problems in Agent Reliability
6. Principles for the Next Era of Agent Engineering

Conclusion

Glossary of Key Terms

References