Agents in Production. Building, Tracing, and Shipping Multi-Step AI You Can Trust
Gabriel Anhaia
Synopsis "Agents in Production. Building, Tracing, and Shipping Multi-Step AI You Can Trust"
At 2:04 a.m. your support agent went into a loop. It called the refund tool, called it again, called it again, and by the time the on-call woke up it had burned through a month of API budget and apologized to the same customer four hundred and twelve times.
An LLM agent is an LLM that decides its own next step in a loop. That one property - the model controls the control flow - breaks every assumption software engineers have about how code behaves. Agents do not just fail the way LLMs fail. They fail in new ways that compound over steps: runaway loops, cost explosions, tool-call misfires, context bloat, state desync, prompt injection through retrieved documents, and silent drift in decision quality.
If you ship LLM features to production - as a backend engineer, a platform engineer, an SRE who inherited someone else's agent - this book is the operational handbook you need. It is not an ML theory book. It is not a prompt engineering tour. It is the stack you actually run on Monday morning to build an agent, trust it, and recover when it breaks.
What you get: agents built first from bare SDKs so you understand the loop, then layered with the frameworks that matter in 2026 - LangGraph, OpenAI Agents SDK, Anthropic Claude Agent SDK, Microsoft agent-framework, CrewAI, and Pydantic AI - each with an honest verdict on what it is good at, what it is bad at, and when to pick it.
You will learn how to trace a full agent trajectory as a tree of spans, how to write evals that score multi-step decisions and not just single outputs, how to apply Meta's Agents Rule of Two and the 2026 security posture against prompt injection and tool abuse, how to deploy agents behind the patterns that keep them cheap and safe, and how to run an incident when your agent is the one on fire. Real code in Python and TypeScript. Real traces. A production-readiness checklist you can run against your own system.
Book 2 of The AI Engineer's Library. Book 1, Observability for LLM Applications, is recommended but not required. Complementary to Hamel Husain's Evals for AI Engineers (O'Reilly, 2026) for eval methodology - this book covers the wider agent lifecycle: building, tracing, guarding, deploying, and recovering.
