Agents and Workflows is 14.7% of the CCDV-F exam and Tools and MCPs is another 10.6%, so a quarter of the paper lives in this course. It covers the four different Anthropic products that all claim to build agents and refuses to let you confuse them, the tool loop at wire level, tool_choice and the strict-tool-use placement trap, the built-in tool catalogue with its real prices and beta headers, Agent Skills including a version conflict the docs have not resolved, and MCP from the two-part connector requirement down to what a server actually exposes. It is honest where the documentation is ambiguous, because the exam is timed and knowing which facts are lookups is worth as much as knowing the facts.
You are building a research assistant. It reads a question, searches the web, runs a bit of code on the results, and writes a summary. Nothing exotic. The point is that every decision you make building it maps onto an exam objective.
Do it in two halves, and the first half is the one that matters.
Write a single markdown file that answers these, with a sentence of justification each. No code, no network, no account. This is the barn wifi half.
input_schema.strict go anywhere? If yes, say exactly which object it goes on. If a tool is coming from MCP, say what that changes.If step 6 kills a choice you made in step 1, that is the capstone working correctly.
Now run it with no SDK loop helper. Write the request body as JSON in a scratch file, then hand-simulate three turns:
stop_reason set to whatever it should be.stop_reason that tells your loop to stop.Then break it on purpose. Write the version where a tool throws, and show the exact shape you send back so the model can recover instead of your process crashing.
You are done when someone could read your file and implement the thing without asking you a single clarifying question, and when the failure turn is in there. The failure turn is the part everybody skips, and it is the part the exam asks about.