Weight: 19%, the largest domain. About 12 of the 63 exam questions.
This domain is about connecting Claude to tools, data and other systems safely, quickly and observably.
| Objective | What it tests |
|---|---|
| 3.1 Capability bloat | Too many or overlapping tools, least privilege, per-role tool sets |
| 3.2 Authentication and authorization | Per-user access, tenant isolation, scopes, secrets, confused deputy |
| 3.3 Accuracy-latency trade-offs | Justifying configurations with measured data against SLAs |
| 3.4 Observability at scale | Tracing, sampling, cost attribution, SLO alerting |
| 3.5 RAG chunking and indexing | Structure-aware chunks, contextual chunks, parent-child, re-indexing |
| 3.6 Retrieval strategies | Hybrid, metadata filters, SQL tools, agentic and multi-hop retrieval |
| 3.7 Integration mechanisms | MCP vs direct API/CLI vs tool use vs agent-to-agent |
| 3.8 Progressive discovery vs monolithic context | Loading tools and instructions on demand |
Mental model
- The model decides; your systems enforce. Authorization never depends on the prompt or on values the model generates.
- Every tool is context cost and attack surface.