Vector embedding, fine tuning and task-specific language models are examinable topics on this credential, sitting in its heaviest domain. Most study material still describes the syllabus that came before them.
Fitting a model accounts for just over half this paper. The rest is the work either side of the fit, and those are the domains an analyst is least likely to have picked up informally. Here is how the five domains divide, and why input preparation and model measurement should be studied as one subject.
Native Apps is the SnowPro Specialty exam that treats you as a software vendor rather than a data professional. Manifests, setup scripts, execution rights, release channels and Marketplace listings carry the paper, and two thirds of the marks sit in the two domains that assume you have actually shipped something.
The capabilities are the same and the labels are not. Google has confirmed that this exam was updated for the move from Vertex AI to Gemini Enterprise Agent Platform, which means the component names in last year’s study notes are the wrong answers now.
On most protocol exams architecture is the biggest domain. On MCPA it is the smallest at 14 percent, while running a tool call and authorising it take half the marks between them. Sixty questions, 120 minutes, 75 percent, and a specification published as dated revisions rather than numbered releases.
Sixty correct answers from seventy-five, across six domains weighted within five points of each other. There is no topic you can write off, and the biggest one is retrieval rather than prompting.
Most blueprints leave depth to the reader. This one lists CATX, SCAN, MDY and OBS= by name, which turns preparation into a checklist you can genuinely finish.