A hands-on companion for engineers who ship LLM agents and cannot currently say how often theirs is wrong.

52 pages, five modules, in build order. Each one produces the artefact the next consumes, so the sequence is not decorative.

Build a support agent with a real spec and permissions enforced in code, not in the prompt. Instrument it before it takes traffic, with a trace schema you will not have to rewrite. Generate a deterministic synthetic world you can break safely.

Then the part most teams skip: read the traces yourself. Open coding, axial coding, and the research finding that explains why you cannot write your quality rubric before you have graded anything. Build one binary evaluator per failure mode, validate it against human labels, and report both its error rates rather than its accuracy.

Turn those failures into a regression gate. Learn why capability and reliability are different questions, and why the second is the one your users experience. Attack your own endpoint and convert every successful attack into a test that blocks forever. Then compare configurations honestly on accuracy and cost, and keep more than one.

Seven worksheets you complete against your own agent. Two vector diagrams. A glossary, interview traps, and an honest page on which claims here you should go and verify yourself.

Print-ready A4 PDF.