Agentic AI in Production: Ship, Scale, and Govern Autonomous Systems - Governance· Identity· Operations· Testing

Agentic AI in Production is the book for engineers who have crossed the gap from "agent demo" to "agent serving traffic." Itcovers the production concerns most agent tutorials skip: enterprise identity, secrets, governance, observability, and the operational discipline required to run autonomous systems where mistakes have consequences.What you'll learnEnterprise identity for agents — X.509 certificates, OIDC delegation, zero-trust patterns, and the difference between authenticating a human and authenticating an agent acting on a human's behalfSecret management — HashiCorp Vault, AWS Secrets Manager, Azure Key Vault compared with a clear decision framework, plus rotation, JIT injection, and the bank-vault vs hotel-safe mental model for picking a backendHierarchical rate limits and budgets — token-bucket math derived from first principles, the cost-cascade closed form that explains why retry storms get expensive fast, and graceful degradation patterns when budgets biteTamper-evident audit trails — hash-chained event logs with PII tokenization, integrity verification, and the precise security guarantees you can and cannot claim about a chain anchored to external checkpointsDeployment strategies — serverless, containers, VMs, and Kubernetes compared by blast-radius semantics; blue-green and canary patterns with the statistical math (two-proportion z-test) for sizing real canary trafficMonitoring that fits agent workloads — USE/RED/Four Golden Signals adapted for LLM latency tails, anomaly detection that does not over-alert on heavy-tailed distributions, and the percentile estimator you actually want.Error handling and resilience — retries with full and decorrelated jitter, circuit breakers with the math, fallback chains, dead-letter queues, and the geometric expectation formula that tells you when retry is hopelessTesting agent systems — a six-layer pyramid extending classical testing, MockLLM patterns for unit tests, behavioral evaluators, and the binomial math for picking a stability threshold.What makes this book differentProduction-aware code, not toy snippets. Companion code at github.com/vijaygwu/Ship-Scale-and-Govern-Autonomous-Systems ships with bounded collections, explicit timeouts, retries with jitter, circuit breakers, dead-letter queues, and AGENT_ENV=production trip-wires that refuse to deploy demo material. Every chapter module has matching regression tests.Math you can actually use. Cost cascades, canary sample sizes, Little's Law sizing, geometric retry expectations, binomial flakiness thresholds derived from first principles with worked examples, not buried in citations.Honest hedging. Composite case studies are labeled. Anecdotal figures are flagged. "Tamper-evident" is not promised without the external anchor. "Always" and "guarantees" appear only where the math supports them.Who this book is forEngineers shipping LLM-backed agents into environments where mistakes have audit consequences: financial services, healthcare, enterprise SaaS, customer-facing applications with compliance requirements. Familiarity with Python and basic LLM concepts assumed; previous reading of Agentic AI (Book 1) helps but is not required.606 pages. Eight chapters. Hundreds of runnable code samples. Production discipline distilled. Read more