Harness Engineering: Building Reliable AI Agents That Actually Deliver

Everyone in the room saw the agent work. That is the problem. In production it fails one run in twelve, and you carry the pager.You are not stuck because the agent cannot do the work. You watched it do the work. You are stuck because doing it once, on stage, with you watching is the easy part, and doing it every time, while you sleep, is the job nobody named for you. This book names it: the demo cliff. Harness Engineering is the climb back up.You already built an agent. This book is about making it deliver. Evals, verification, guardrails, observability, recovery: the scaffolding that turns an impressive toy into a system you would put your name on the pager for.A reliable agent still fails. Its failures are rare, cheap, caught before the user, and survivable when they are not. That is engineering, and this book is the engineering.What you build- Reliability as a number: a success rate against a fixed task set, tracked run over run.- An eval gate that blocks a bad change before it ships, run like CI.- A verification wall that catches wrong output before a user does, cheap checks first, expensive ones behind them.- A recovery path for the failures you cannot prevent: retry with judgment, fall back, escalate, roll back.- Run records of what the agent actually did, so a bad run is a query instead of an archaeology dig.- A reliability budget for one real agent: a target you can defend and the guardrails that hold it.Who it is forSoftware, ML, and platform engineers who already run agentic loops and now own one in production. You shipped the feature, the demo recording is still in the channel, and the 3 AM page lands on you. If that is your week, this is your book.About the authorWes Halloran writes practical books for working developers and engineering leaders in the shift to AI-agent development. He has shipped agents that failed in ways no demo hinted at, so this book trusts measured runs over impressive ones. He works from real tool behavior and verified workflows, and never accepts a green light he did not check.Part of the AI and Agentic Engineering series.

Pages: 170, Paperback, Independently published