Weight: 18%. About 11 of the 60 questions.
| Topic | What it tests |
|---|---|
| 2.1 Tool interfaces | Descriptions, names, boundaries, input and output design |
| 2.2 Structured errors | isError, error categories, isRetryable, actionable messages |
2.3 Tool distribution & tool_choice | Least privilege, tools per agent, auto / any / forced / none |
| 2.4 MCP integration | Scopes (local / project / user), .mcp.json, env vars, resources, prompts |
| 2.5 Built-in tools | Read, Write, Edit, Bash, Grep, Glob, and when to use each |
Mental model
- The tool description is the model's main signal for choosing a tool. Fix confusion at the interface first.
- Errors are information for the model: say what failed, whether to retry, and what to do instead.
- Fewer, focused tools per agent means more reliable selection.
Exam heuristics
- Misrouting between similar tools: improve the descriptions and boundaries, or split or rename the tools. Check the system prompt for keyword bias.
- An agent gives up, or retries pointlessly: the error response lacks a category, retryability, or a remedy.
- Team-shared MCP config goes in project
.mcp.jsonwith${ENV_VAR}secrets. Personal config goes in user or local scope.
Updated 2026-10-28