Marke: Javier Rayón
Variante: Taschenbuch
Eigenschaften:
Second volume of the AI Engineering series, from learnaipress.com. Four books covering the four layers of working with coding agents: the loop that sustains the cycle, the harness that makes it reliable, the graph that splits work across specialists, and the specification that decides what gets built. Each one reads on its own.This one covers the second layer. A well-directed agent writes reasonable software, and the next day repeats the mistake you fixed by hand yesterday, because the fix stayed in a conversation nobody is going to re-read. The harness is the environment around the agent — guides, sensors, limits and telemetry — that turns each observed failure into a permanent piece. This book teaches you to build it, operate it and share it with a team.Guides and sensors: the vocabulary that orders the workGuides act before the agent writes anything: the project memory file, the templates, the scaffolding, the permission profile. Sensors act afterwards: the tests, the type-checker, custom linters, continuous integration. A second axis crosses that one, the nature of the control: the computational runs on CPU, is deterministic and cheap; the inferential is resolved by a model, costs more, and reaches a semantic judgement no linter reaches.With those four boxes in front of you, you can decide what to build first — the question a team cannot answer when it says "we need a better harness".What you will learnThe rule everything hangs from: each agent failure becomes the piece that prevents the next one.Preparing a repository so it can be governed: strict typing, module boundaries and a reproducible build.Writing a memory file where every line exists because of a real failure, and pruning it when it expires.Tests as a self-correction loop, with mutation testing against the suite that approves itself.Evals: your agent's own regression suite, with cases taken from your failures and a checkable criterion.Permissions, sandboxing and blast radius, sized by the cost of the worst possible error.Blocking hooks with the real PreToolUse contract, and where a permission list leaks.Context measured in tokens, harness observability, and the steering loop that decides which control gets raised.When the person writing the prompt is not youAn entire part deals with the shared harness: whose memory is whose and with what precedence, who may change the agent's configuration, what happens to a new hire on day one, and what happens when someone bypasses the local controls. It closes with the text that enters the model's window without anyone on your team having written it — the body of an issue, a dependency's README, the response from a connected server — and the configuration rule that bounds it.Who it is forDevelopers running agents in production who want a hit rate that does not depend on the prompt of the day.Tech leads who authorise long unsupervised tasks and need the damage bounded in advance.Teams with uneven experience who want code quality that does not depend on who launched the task.About the toolsThe examples use Claude Code, with concepts portable to any agent: guides, sensors, permissions and external verification transfer by changing the commands, not the design. It takes the individual pieces as known — permissions, hooks, MCP, subagents — and teaches you to compose them into a control system.Fifteen chapters and an appendix of copy-and-paste material. You pick a real repository in the first chapter and build its harness through to the last. The paperback edition is printed in full color.
