The first certification for a discipline most teams are doing badly.Your agent worked in the demo. It degrades in production, and the explanations you reach for - the model, the prompt, bad luck - are all wrong in the same way. They look at what the agent did rather than at what it was given.A context window is finite. Everything admitted to it displaces something else. Almost every failure blamed on capability is a decision about that window that nobody made deliberately.Databricks launched the industry's first certification for this on 29 July 2026. This guide is written against its published exam guide, section by section, with every chapter traceable to a stated objective.An exam that asks what an agent can afford to knowLook at how the official sample questions end. Without discarding previously retrieved facts. Most cost-effectively. Without significantly increasing token cost. Without expanding each sub-agent's context window. With the least maintenance burden.Five sample questions, five closing clauses, each one eliminating a design that would otherwise work. Not one of them asks what an agent could usefully know. All of them ask what it can afford to. This book is organised around that fact.Why experienced agent builders lose marksYour agent works under conditions you never had to notice. The sessions are short, because you are the one testing it. The corpus is curated. The tools are yours, so none of them overlap confusingly. Nobody has asked what it costs per thousand conversations, because it has not had a thousand.The exam supplies every condition your demonstration never did - and the design you would build appears among the options precisely because it is what most engineers would build.What is insideEleven chapters - one per published exam section, plus reasoning foundations and exam strategyTwo complete practice exams - 90 questions total, matching the real paper's length, with every answer and the reason each distractor fails sitting directly beneath its question88 review questions across the chapters, every wrong option explainedSeven Design Challenges - multi-constraint scenarios resolved by decision before anything is builtThree Context Budget Drills - a token ceiling, more candidate content than fits, and an allocation you must commit toEleven Trace Reading Blocks - diagnose from what actually entered the window, which is what several objectives ask for by nameA 72-entry control reference - every technique with what it does, what it cannot do, what it is confused with, and the fact that separates themCovered in fullThe four context failure modes - poisoning, distraction, confusion, clash - and how to tell them apart from a trace. System prompts and the instructions that cannot work at any wording. Databricks AI Search, index synchronisation, chunking, and Genie space configuration. Memory architecture on Lakebase, and why over-retrieval is the failure that never generates a complaint. Tool design, MCP, progressive disclosure, and Agent Skills. Compaction, and why tuning it for brevity first cannot be evaluated. Multi-agent propagation and orchestrator saturation.Who this is forEngineers who have built something with a language model and watched it behave differently at turn fifty than at turn five. Six months of hands-on experience is about right. You do not need to have built a multi-agent system - but you should have added a tool that made something else worse, or retrieved a corpus and wondered why the answers got vaguer.What this is notNo screenshots, no click-by-click walkthroughs. The exam never asks you to perform a configuration; it asks which configuration a stated constraint demands.
