DOMAIN-SPECIFIC HARNESSES

CHECK YOUR UNDERSTANDING • When should a support agent escalate to a human? ◦ Answer: When the issue is complex or risky, the customer is upset, a legal or safety concern appears, or the agent has low confidence. • Why show the query with a data analysis answer? ◦ Answer: A plausible query can be wrong, and showing it lets an analyst check what was actually counted. • What is the rule of thumb for regulated fields? ◦ Answer: The agent drafts and a qualified human decides when the consequences for a person are serious. • Name two advantages of an edge agent. ◦ Answer: Any two of speed, privacy, and reliability without a network connection. • What does the same set of design questions decide? ◦ Answer: The permission levels and verification steps suited to the domain's risks. PART SIXTEEN: BUILDING A HARNESS FROM SCRATCH (CAPSTONE PROJECTS) Fifteen parts of this book explained how a harness works. Part Sixteen asks you to build one with your own hands. You will build a small agent called the Resume Review Agent, and you will build it in five stages, one stage per chapter. Each stage adds a new layer of safety or skill, in the same order in which the book taught the ideas. By the end, the little agent will touch all seven layers of our organizing map, which we call the ETCLOVG taxonomy, a name made from the first letters of Execution, Tooling, Context, Lifecycle, Observability, Verification, and Governance. Along the way you will meet bugs that we planted on purpose, and you will find and fix them, which is the best training a harness engineer can get. Everything in this part is fictional and safe, so you can experiment without fear of hurting anything.