CCAR-P Learning Guide
Domain 3 · 19% of the exam

Integration

Connecting Claude to enterprise systems safely: least-privilege tool sets, authentication and authorization, observability at scale, RAG pipeline design, and choosing between MCP, API/CLI and cloud delivery routes.

  1. 3.1Choosing an Integration MechanismChoose between the API, the SDKs, the Agent SDK, MCP, Claude Code and agent-to-agent designs by what Claude has to do and who uses it, and pick the delivery route the organization's constraints allow.
  2. 3.2Authentication, Authorization and Security GapsEstablish identity on the server, enforce authorization in the systems that own the data, give tools the least privilege they need, and spot the gaps where a design trusts the model or the user instead.
  3. 3.3Avoiding Capability BloatRecognize when an agent has more tools, servers or permissions than its task needs, and cut it back to a small, well-described tool set that is cheaper, more accurate and safer.
  4. 3.4Progressive Discovery versus Monolithic ContextDecide whether an agent should get everything up front or discover tools, knowledge and instructions as it goes, and use Skills, tool search and just-in-time retrieval to keep a large capability set from flooding the context.
  5. 3.5Designing a RAG PipelineDesign a retrieval pipeline end to end, choosing chunking by the shape of the documents and indexing by how people ask, and keep the index fresh and measured.
  6. 3.6Matching Retrieval to Data and QueriesMatch the retrieval method to the shape of the data and the way questions are asked, from vector search and keyword search to structured queries, metadata filters, reranking and agentic search.
  7. 3.7Accuracy-Latency Trade-offsKnow which configuration settings move accuracy and which move latency, find the setting that holds up the tightest budget, and justify each choice with measurements rather than preference.
  8. 3.8Observability at ScaleInstrument Claude integrations so any answer can be traced end to end, costs and limits can be attributed to the right team or tenant, and problems are detected before users report them.